Work
Production-minded builds spanning AI agents, full-stack SaaS, search systems, and cloud-delivered products.
AI Research Platform
ICISML 2026 Published · Enterprise LLM Research Orchestration Platform
Enterprise-grade collaborative research platform with LLM orchestration, multi-agent workflows, real-time collaboration, and AI-driven paper recommendations. Research paper on this platform was submitted to and accepted at ICISML 2026 (International Conference on Intelligent Systems, Machine Learning) — conference completed.
AI Search & RAG
Built SaaS Product in 2 Days with Claude Code.
EdTech SaaS platform generating curriculum-aligned MCQs from textbook PDFs using RAG and GPT-4o. Students select a chapter and receive a 10-question test with difficulty-distributed explanations in seconds.
AI Agent
Autonomous Competitive Intelligence • SWOT & Positioning Analysis
Autonomous multi-agent system that researches and analyzes competitive landscapes in real time. Agents crawl company data, news, and market signals to generate structured competitor profiles, SWOT analyses, positioning maps, and strategic recommendations — delivered as interactive reports.
AI Platform
Built with Claude Code under 6hrs.
Master-level platform for system design interviews. Features privacy-first local LLM integration via Ollama, RAG-powered assistants, and automated ATS resume analysis.
AI Research Platform
AI-Powered Research Hub • LangGraph Chatbot • Multi-Agent Workflows
A comprehensive research ecosystem featuring a LangGraph-powered RAG chatbot, multi-agent research orchestration, and real-time collaborative tools. Includes AI-driven paper recommendations, peer review workflows, and automated citation management.
AI Agent
Automated AI Career Agent • Built with Antigravity under 10hrs.
Automated job crawler and matching platform. Features intelligent resume matching, company insights, and real-time data integration to connect seekers with ideal roles.
AI Platform
Gamified Interview Prep • AI Coaching & XP Progression
Gamified daily interview preparation platform delivering curated questions with AI coaching (GPT-4o-mini), XP/leveling system (Beginner→Elite), daily streaks, topic mastery tracking, and achievement missions to build sharper interview skills through consistent practice.
Web Development
Premium Design • AI Chat • High-Performance Framework
The 6th iteration of my personal portfolio, featuring a premium redesign focused on a modern aesthetic with a high-performance technical foundation. Includes an AI-powered chat widget, serverless backend, and advanced SEO optimization.
Web Development
GitHub Profile → AI Story Page • GPT-4o Narrative • Shareable URLs
A full-stack app that turns any public GitHub profile into a beautiful, shareable developer story page. Users enter a GitHub username, the Express backend fetches public profile data, OpenAI GPT-4o generates a narrative, and the React frontend renders an interactive career story at a unique shareable URL — with Replit DB caching, Open Graph meta tags, and a dark UI with explore page.
Infrastructure
9-Provider API Key Validator • Server-Side Security • Stress Testing
Professional dashboard for validating and stress-testing API keys across 9 AI providers (OpenAI, Anthropic, Gemini, Groq, Mistral, Perplexity, OpenRouter, DeepSeek, Together AI). All validation runs server-side via Next.js Server Actions — keys never leave the browser. Features model stress testing with latency benchmarks, searchable audit logs, JSON export, and 15-second timeout protection.
Deep Learning
LSTM-CNN-GRU Ensemble • 92% Prediction Accuracy
Hybrid deep learning system (LSTM-CNN-GRU) for real-time traffic forecasting. Features Graph Neural Networks for spatial city-wide analysis and interactive Google Maps routing.
Web Development
Multi-Tenant Sports SaaS • Real-time Scoring & Leaderboards
Production-ready multi-tenant SaaS platform for managing badminton matches across teams. Features real-time scoring with official rules (first to 21, best of 3), AI-powered doubles pairing, player leaderboards, venue management, and mobile-first courtside UX.
AI Platform
Decode your Developer DNA from your GitHub profile
AI-powered analysis that scans your GitHub repositories, languages, and contribution patterns to reveal your unique developer identity — visualized as an interactive 3D DNA helix.
AI Platform
AI-Powered Stock Forecasting • Live Streaming • Paper Trading
A local-first, SaaS-style stock analytics platform with AI-powered forecasting, live streaming, sentiment analysis, and paper-trading workflows. Features auto-selection of ML models (LSTM, TFT, Stat) and a robust Python/Flask backend.
AI Agent
Conversational Voice Interface • Low-Latency Streaming • Function Calling
A Next.js 14 starter kit for building conversational voice agents in the browser. Features voice input with Web Speech API, live waveform visualization, streaming LLM responses for low latency, and optional ElevenLabs TTS integration. Supports function calling tools (weather, web search, math) and persists conversation memory across sessions.
AI Agent
8-Agent A2A Pipeline • Startup Blueprint in Minutes
Autonomous multi-agent system that takes a single startup idea and produces a complete startup blueprint — market research, competitor analysis, PRD, system architecture, MVP code, go-to-market strategy, and investor pitch deck. Built on Google's A2A protocol with 8 specialized agents (Perplexity, Claude, GPT-4.1, Gemini) and real-time SSE streaming.
AI Agent
Autonomous Infrastructure • Multi-Agent Orchestration • Self-Healing CI/CD
A comprehensive multi-agent system for autonomous infrastructure management, log analysis, and CI/CD optimization. Features specialized agents for remediation, orchestrating complex workflows with zero cloud dependencies using local LLMs.
AI Agent
Production AI Safety Guardrails • 12 Configurable Rules • Prometheus Observability
Python framework that wraps any AI agent with production-grade safety guardrails — compatible with LangChain, LangGraph, CrewAI, or custom callables. Features prompt injection detection, PII/secrets filtering, cost budget controls, hallucination reduction, human escalation workflows, and full observability via Prometheus metrics and a Streamlit dashboard.
AI Search & RAG
Agentic Retrieve → Evaluate → Refine Loop • Multi-Source RAG
Intelligent RAG system where an LLM agent autonomously decides when retrieval is needed, what to query, which sources to consult, and when to re-query. Implements a retrieve → evaluate → refine → re-retrieve production loop with multi-source retrieval (vector, web, SQL), relevance evaluation, and optional fact-checking — orchestrated by LangGraph.
AI Search & RAG
Multimodal RAG • Knowledge Bases • Citation-Backed Chat
Multimodal RAG application for uploading documents and images into knowledge bases and chatting with grounded, citation-backed answers. Features knowledge-base organization, OCR and semantic extraction for image uploads, Celery-based async ingestion, pgvector semantic search, persistent conversation history, and a Next.js dashboard with authentication and upload workflows — all packaged as a Docker-based local stack.
AI Search & RAG
RAG Debugging Workspace • BM25 vs Vector • Sentence-Level Attribution
Local-first full-stack debugging workspace for retrieval-augmented generation systems. Runs queries against a bundled corpus, compares BM25 and vector retrieval side-by-side, builds extractive grounded answers with sentence-level attribution back to supporting chunks, assigns a hallucination-risk heuristic based on sentence support, and tracks every run in SQLite — visualized as a debugging dashboard with retriever panels, hallucination signals, and an experiment history stream. No external LLM API key required.
AI Search & RAG
Multimodal PDF RAG • GPT-4o Vision • Hybrid Retrieval
Production-ready RAG pipeline processing PDFs with text, images, tables, and charts. Uses GPT-4o Vision for visual understanding, generates semantic descriptions for indexing, and provides unified hybrid retrieval across all content modalities with page-level source attribution.
AI Agent
5-Agent PR Review • LangGraph Orchestration • Security Scanning
Multi-agent PR review system deploying 5 specialized AI agents in parallel: code quality, security vulnerability scanning, performance bottlenecks, test coverage, and an orchestrator that synthesizes results. Features LangGraph orchestration, SHA-256 caching, and GitHub PR commenting.
AI Agent
5 Concurrent Review Agents • Git Diffs & GitHub PRs • Multi-Backend
Multi-agent code review CLI for Git diffs and GitHub pull requests. Runs five specialist LLM agents — security, performance, style, tests, and docs — concurrently over a unified diff via asyncio, then merges their schema-validated findings into a single structured review: deduplicated by overlap, ranked by severity, capped per category, and rendered to the terminal, Markdown, JSON, or posted back to a GitHub PR as an inline review. Includes a FastAPI webhook server with HMAC verification for auto-reviewing PRs and three backends (OpenAI, local Ollama, zero-dependency mock).
AI Platform
Automatic Prompt Optimization • Instruction Search + Few-Shot Bootstrapping • DSPy-Inspired
Automatic prompt optimization for LLM tasks, inspired by DSPy. Declare your task as a Signature with a few labeled examples and PromptForge searches for the prompt that maximizes your metric — combining InstructionSearch (LLM-proposed instruction rewrites, hill-climbed on a validation split) with BootstrapFewShot (mining high-scoring demos as in-context examples). Every candidate is scored on held-out data and logged to a SQLite leaderboard, and all model calls are content-hashed and disk-cached so re-runs are free. Runs on OpenAI or fully local Ollama models.
AI Agent
Local-First Terminal Coding Agent • ReAct Tool-Use Loop • Ollama-Powered
Local, terminal-based coding agent that runs entirely on your machine via Ollama — a "mini Claude Code" that understands a repository, plans changes, edits files, runs commands and tests, and iterates autonomously with zero cloud dependency by default. Built on a ReAct tool-use loop where the model reasons step-by-step and calls tools (read/write files, apply patches, run commands, grep, undo) under a strict JSON protocol with defensive parsing and self-correcting error feedback. Uses tree-sitter symbol mapping and semantic retrieval to keep only relevant files in context, shows unified diffs before applying changes, sandboxes commands to the project directory, and offers one-step undo. Ships both a Rich REPL and a Textual TUI with transcript and plan panes; core Ollama communication uses only the standard library, with tree-sitter and embeddings as optional pure-Python-fallback extras.
AI Platform
Prompt Compression Toolkit • Rule-Based + LLMLingua + Abstractive • Token & Cost Savings
Python toolkit for compressing long LLM prompts — system prompts, RAG contexts, and code explanations — to cut token usage and cost while preserving the information that matters. Offers three pluggable compression strategies: rule-based (local and fast), an LLMLingua-like model-based token scorer, and abstractive LLM rewriting, all behind a simple Compressor API with configurable compression ratios. Includes drop-in wrapper clients for OpenAI and Anthropic and a `psq` CLI, plus a benchmark framework with bundled datasets and HTML reporting across three tasks — long summarization (Rouge-L), RAG question-answering (exact match), and code explanation (BLEU).
AI Agent
First Open-Source A2A Protocol • 6 Specialized Agents
One of the first open-source implementations of Google's Agent-to-Agent (A2A) protocol. A marketplace where specialized AI agents discover and collaborate on complex tasks via natural language decomposition. Features 6 agents (code, content, data, research, translation, design), dependency-aware execution, and real-time SSE progress tracking.
AI Agent
Multi-Step Research Agent • LangGraph + SSE Streaming
Autonomous research assistant with FastAPI backend and Next.js frontend. Processes research questions through multi-step LangGraph agent workflows, delivering markdown reports, extracted findings, credibility-scored sources, and quality critiques — all streamed live via Server-Sent Events.
AI Security
Full-Stack AI Red Team Platform
A full-stack AI red team platform for testing LLM applications against prompt injection, jailbreaks, data exfiltration, and tool misuse. Features a plugin-based attack engine and async execution pipeline.
AI Security
LLM Security Engine • Rule-based + LLM Detection
A robust detection engine safeguarding LLM applications against prompt injections, jailbreaks, and system prompt leaks. Features rule-based heuristics, statistical analysis, and LLM-powered classification, complete with an interactive Next.js dashboard.
AI Platform
Adversarial LLM Debates • Round-by-Round Scoring • Batch Evaluation CLI
A production-ready Next.js app for running adversarial debates between two LLMs, scoring each round with a judge model, saving debate history, and exporting full transcripts. Features live streamed debates, third-model judge scoring, and a CLI evaluation mode for batch experiments.
AI Search & RAG
Lexical RAG (No Vector Store) • BM25 + TF-IDF Fusion • Grounded Citations
A production-style Retrieval-Augmented Generation (RAG) application that answers questions over uploaded documents WITHOUT using a vector database. Implements lexical search (BM25 + TF-IDF) with page-aware and section-aware retrieval, hybrid fusion ranking, and grounded citations for verifiable answers.
AI Agent
Natural Language to SQL • Multi-DB Support • Query Optimization
A full-stack AI database assistant for natural-language analytics. Connects to PostgreSQL, SQLite, and DuckDB to provide schema-aware SQL generation, query plan explanation, and index optimization suggestions with read-only safety guardrails.
AI Agent
Closed-Loop Query Rewrites • Equivalence Checks • Read-Only Safe
A closed-loop SQL optimization assistant for PostgreSQL and DuckDB. Profiles a read-only query, asks an LLM for a safer rewrite and index ideas, reruns the candidate, verifies practical result equivalence via row counts and bounded samples, and reports measured timing differences. Advisory by design — it never applies indexes automatically and refuses non-read-only SQL.
AI Search & RAG
Evidence-Backed Knowledge Graphs • NL-to-Cypher • Neo4j Visualization
Builds an evidence-backed Neo4j knowledge graph from a folder of PDFs, then lets you query it via natural language or hand-written Cypher in the browser. Extracts text with PyMuPDF, asks Claude for structured entities and relations with evidence quotes, resolves duplicate mentions with local embeddings, and serves a FastAPI + vis-network graph explorer.
AI Search & RAG
Cache-Augmented Generation • 50–200ms Latency • 50% Cost Savings vs RAG
Retrieval-free document Q&A system using Cache-Augmented Generation (CAG) as a faster alternative to RAG. Preloads the complete knowledge base into GPT-4o's KV cache, eliminating vector search latency entirely — achieving 50–200ms response times vs RAG's 200–600ms with 50% cost savings on cached tokens.
AI Search & RAG
Knowledge Graph + Vector Search • Entity Relationships • Graph Traversal
Knowledge Graph-Enhanced RAG that combines vector search with graph traversal for deeper question answering. Extracts entities and relationships from documents, builds a NetworkX knowledge graph, and traverses entity connections to discover context that vector-only search misses — with interactive pyvis visualization.
AI Search & RAG
Semantic Code Search • 15+ Languages • Dependency Graphs
RAG system for querying any GitHub repo or local codebase via natural language. Uses Tree-sitter for semantic code chunking (functions/classes), ChromaDB for vector storage, cross-encoder reranking, and dependency graphs for impact analysis. Supports 15+ languages with CLI and REST API interfaces.
AI Search & RAG
Contract Review AI • Hybrid Retrieval • Risk Scoring
AI-powered contract review system with clause extraction, risk scoring, and document-grounded Q&A. Features hybrid retrieval (semantic 70% + BM25 30%), risk classification (Low/Medium/High), streaming chat, and markdown report export across PDF, DOCX, and TXT formats.
AI Agent
Autonomous Gmail Triage • Local Preference Learning • Automated Unsubscribe
AI-powered Gmail inbox triage system that classifies messages, drafts replies, archives low-priority mail, and automates unsubscribe flows. Features local preference learning through a rules engine, Playwright-based automation, and a SQLite audit log for decision tracking.
AI Search & RAG
o1 Reasoning • Agentic Knowledge Discovery
Sophisticated knowledge-base search engine using OpenAI's Agent SDK and o1 models. Features intelligent PDF processing, vector-based semantic search, and reasoning-driven retrieval.
AI Agent
4 Autonomous Agents • Self-Hosted DevOps • Zero Cloud Dependencies
Self-hosted multi-agent system for infrastructure management, CI/CD optimization, and automated diagnostics. Features 4 autonomous agents — Infrastructure (Docker/K8s/Terraform generation), Pipeline (CI/CD analysis & optimization), Heal (CrashLoopBackOff/OOMKilled diagnosis with remediation), and Orchestrator (sequential & parallel workflows). Powered by Ollama (Llama3 + CodeLlama) with FAISS-based RAG. Zero cloud dependencies.
Machine Learning
KAN-Powered Crop Risk • Algorithmic Recourse • Interpretable AI
Interpretable multimodal crop risk prediction platform using Kolmogorov-Arnold Networks (KAN) — telling farmers not just what will happen, but why and what they can do about it. Features native spline-based explainability, constrained algorithmic recourse over mutable features (NPK, soil pH, area), multi-dataset validation across Indian and UK agricultural datasets, and a FastAPI + React dashboard with SHAP, KAN spline charts, and model comparison across XGBoost, CatBoost, RF, MLP, and KAN.
Deep Learning
Multimodal AI • Real-time Classification • XAI Interpretability
A comprehensive multimodal deep learning system for brain tumor classification and analysis. Features a modern React frontend and a robust Flask backend for real-time inference, with advanced diagnostic capabilities including probability distribution visualization, and XAI (Grad-CAM) for model interpretability.
Computer Vision
TDA + Tensor Fusion • 92.45% Accuracy • GradCAM++ Explainability
Research-grade web app for explainable brain tumor classification from MRI scans. Fuses Topological Data Analysis (TDA), Tucker/Tensor Train decomposition, and classical ML — achieving 92.45% accuracy (SVM-RBF) with GradCAM++ heatmaps, TDA-derived ROI masks via persistent homology, and quantitative IoU/Dice explainability metrics across 4 tumor classes.
Machine Learning
Hybrid Recommender • FAISS + SVD • A/B Experimentation
Production-grade recommendation platform demonstrating end-to-end personalization with collaborative filtering, content retrieval, hybrid ranking, explainable recommendations, deterministic A/B/C experimentation, and admin analytics for model operations.
Machine Learning
Local RL Playground • Multi-Algorithm Support • Real-time Monitoring
A local-first reinforcement learning playground for training and comparing RL agents. Supports multi-algorithm training (PPO, DQN, A2C), classic control environments, and custom GridWorlds with real-time WebSocket-based monitoring and visualization.
Deep Learning
Multimodal AI • IoT Integration • Disaster Response
A comprehensive disaster management platform that integrates IoT sensor data, social media analytics, and satellite imagery (xBD dataset) for real-time intelligence and AI-driven resource allocation. Features DeepLabV3+ for damage assessment and real-time social media processing.
AI Automation
Autonomous Testing • RAG-Powered QA Agent
AI-powered test automation system leveraging RAG to generate comprehensive test cases and executable Selenium scripts. Features semantic search with ChromaDB.
AI Platform
Microservices Architecture • Real-time Intelligence
Comprehensive microservices-based AI platform featuring RAG knowledge bases, real-time stock market analysis with Kafka, and autonomous agent workflows.
AI Platform
AI-Powered Hiring • Automated Evaluation
AI-powered recruitment system using Django, Google Gemini, and OpenAI GPT. Features automated technical question generation, intelligent candidate evaluation, and resume parsing.
AI Agent
ATS Scoring • Skill Gap Analysis • Auto-Tailored Resume
AI-powered resume customization tool that analyzes skill gaps between your resume and a job description, rewrites bullet points to match JD keywords, scores your ATS fit, and generates a fully tailored resume — powered by GPT-4o-mini with spaCy NER and scikit-learn keyword extraction.
AI Agent
Semantic Career Matching • Skill Gap Analysis
Intelligent career assistant using semantic skill analysis and GPT-powered matching. Features multi-source job fetching, skill gap identification, and application tracking.
AI Platform
LLM Context Visualizer • Drag-and-Drop Building • Token Optimization
A browser-based devtool for building, reordering, and compressing LLM context windows visually. Features drag-and-drop context blocks, real-time token counts, and automated optimization.
AI Platform
LLM-Powered Data Generation • Quality Assurance • Streamlit Dashboard
A modular framework for generating high-quality synthetic datasets using Large Language Models (LLMs) via OpenRouter. Includes automated quality evaluation, customizable generation templates, and a Streamlit monitoring dashboard for data curation.
AI Platform
Scalable Data Engineering • Full-Stack Monorepo • Next.js + Python
A full-stack monorepo for architecting and managing AI data pipelines. Features a robust Next.js web interface, Prisma ORM for lineage tracking, and Python-based worker services for scalable data ingestion and processing.
AI Platform
Multi-Speaker TTS SaaS • Coqui TTS Engine
Full-stack text-to-speech SaaS platform built with FastAPI and React. Supports multi-speaker scripts using voice tags, optional duration control, and downloadable WAV audio — powered by Coqui TTS with a modern glassmorphism UI.
AI Platform
Multi-Source Ingestion • Anki Export • GPT-4o-mini
Transform documents into study-ready flashcards with AI. Supports PDF/TXT/MD uploads, URL ingestion, and manual text entry. Configurable quantity, question types, difficulty, and focus areas — with an interactive study mode, keyboard shortcuts, and Anki CSV export.
Web Development
AI-Native Interaction
Next-generation AI portfolio featuring intelligent chat assistance. Built with agentic search and dynamic AI-driven interactions for personalized user experiences.
Enterprise Software
Multi-Tenant Analytics • Real-time Data Sync
Multi-tenant analytics dashboard for Shopify merchants. Features secure OAuth integration, real-time webhook processing, and interactive revenue/product insights.
Infrastructure
Interactive Tokenizer Visualizer • 50 Models • 8 Providers
An interactive, client-side developer tool that visualizes exactly how LLMs and embedding models tokenize your text. Supports 50 models across 8 providers, featuring side-by-side comparison and real-time stats without sending any data to a backend.
Deep Learning
Multimodal Alzheimer's Screening • Speech Biomarkers • MRI Inference
A comprehensive Flask-based research platform for multimodal Alzheimer's screening. Features a dual-branch transformer architecture (WavLM + RoBERTa) for speech and transcript analysis, automated MMSE severity estimation, and SHAP-based explainability. Includes extensive pipelines for acoustic, linguistic, and pause-feature extraction, alongside experimental MRI image inference modules.
AI Agent
Local-First macOS Agent • Ollama + AppleScript Automation • Privacy-First
A local-first macOS desktop agent that runs AI entirely on your machine via Ollama (Gemma 3, Llama, Mistral) with optional cloud fallback (OpenAI, Anthropic). Features macOS automation via AppleScript — opening apps, reading files, inspecting Finder, drafting emails — plus screenshot capture with Tesseract OCR, a risk-leveled permission system, global Cmd+Shift+Space shortcut for instant text analysis, dry-run mode, and a full React dashboard served at localhost:8000 with chat console, action logs, and settings.
Infrastructure
LLM Observability Platform • ClickHouse Analytics • Hallucination Detection
Production-grade observability platform for LLM-powered applications. Monitor throughput, latency, cost, quality, prompt drift, and hallucination signals across your entire AI estate from a single dashboard. Features a live KPI dashboard backed by ClickHouse rollups, a trace explorer for drilling into individual LLM requests, a threshold-based alert engine with webhook delivery, prompt governance with drift scoring, a hybrid hallucination detection evaluator pipeline (lexical heuristics + LLM judge via OpenRouter), a Python SDK with sync/async instrumentation, real-time SSE updates via Redis pub/sub, and pre-provisioned Grafana dashboards.
Infrastructure
OpenAI-Compatible LLM Gateway • Semantic Caching • Rate Limiting & Cost Tracking
OpenAI-compatible proxy that drops in as a self-hosted middleware layer for LLM requests — point your SDK at a new base_url and route every call through the gateway with no code changes. Implements a multi-stage request pipeline (auth → rate limiting → semantic cache → backend routing → cost logging) with FAISS + sentence-transformers semantic caching (~200× faster retrieval at ~4ms vs ~820ms), Redis token-bucket rate limiting via atomic Lua scripts, ordered fallback routing with retries and exponential backoff, per-request token/USD cost tracking, virtual API keys with scoped limits, PII redaction in logs, and a Next.js admin dashboard for usage visualization — all deployable with one Docker command.
AI Agent
Autonomous AI Coding Agent • Docker-Based Validation • Self-Healing Repair Loops
Local-first autonomous AI coding agent platform that transforms natural-language tasks into production-ready code. It features a sophisticated multi-agent workflow orchestrated by LangGraph, isolated execution via Git worktrees, and a secure Docker-based validation sandbox. The system handles the entire development lifecycle: repository scanning, symbol extraction with Tree-sitter, automated planning, multi-file editing with self-healing repair loops, and live progress streaming via SSE—complete with diff approvals and commit artifacts.
Infrastructure
Local-First LLM Observability • Semantic Call Search • PII Redaction & Cost Tracking
Local-first observability platform for LLM-powered applications. A Python SDK wraps OpenAI, Anthropic, Ollama, and custom calls to capture latency, tokens, cost, errors, and metadata — backed by local SQLite storage with export. Features ChromaDB-powered semantic search over captured calls, regex-based PII redaction with allowlisting, model-pricing cost estimation, embedding-based drift detection, OpenTelemetry span exports, and a Next.js dashboard for browsing, comparing, and analyzing requests.
AI Platform
Schema-Driven Document Extraction • Validated JSON/CSV • Provenance & Self-Repair
Schema-driven document extraction tool that converts PDFs and scanned images into structured, validated JSON/CSV. Users upload a document and select or define a Pydantic schema; the system constrains LLM output to that shape, validates results, and runs automated self-repair retries (up to 3 attempts). Features hybrid processing for native PDFs and scanned images with automatic Tesseract OCR fallback, per-field provenance (confidence scores, source text, page numbers, bounding boxes), and support for both OpenAI and local Ollama models.
Infrastructure
OpenTelemetry LLM Observability • Span Waterfalls • Cost, Latency & Eval Scoring
OpenTelemetry-based observability platform for LLM applications that traces model calls, retrievals, and tool invocations into hierarchical span trees — surfacing tokens, cost, latency, errors, and evaluation scores. A Python SDK adds tracing with a one-line @trace decorator and auto-instruments the OpenAI client; a FastAPI collector ingests OTLP/JSON into DuckDB with query and analytics APIs; and a Next.js dashboard renders span waterfall trees and cost/latency charts. Model-agnostic across OpenAI and Ollama, with an offline mock backend for zero-dependency demos and an LLM-as-judge evaluation scoring system.
Infrastructure
End-to-End MLOps Lifecycle • PSI Drift Monitoring • MLflow Tracking & CI
End-to-end MLOps workflow for telecom customer churn that covers the full lifecycle — data generation, model training, experiment tracking, drift monitoring, API serving, testing, and CI — with zero external infrastructure. A unified predict_one() scoring function is shared across the FastAPI service, CLI, and tests to prevent train/serve skew; PSI-based drift detection flags distribution shifts against simulated live data; and optional MLflow integration adds full experiment tracking with no code changes. Reproducible via fixed seeds and shipped with Docker Compose and a GitHub Actions CI pipeline.
AI Platform
100% Client-Side ML • WebGPU Accelerated • Zero Server Calls
Production-grade, browser-based machine learning inference playground that runs text generation, image classification, and text embeddings 100% client-side using WebGPU and Web Workers. Zero backend, zero API keys — your data never leaves your device. Features streaming token output with log probabilities, drag-and-drop image classification with webcam capture, NxN cosine similarity heatmaps with 2D PCA scatter plots, and a configurable benchmark suite with JSON/Markdown export.
AI Agent
Local-First AI Assistant • Tool Orchestration • Cinematic Three.js HUD
A local-first AI assistant with a FastAPI backend, React + Vite frontend, WebSocket streaming, SQLite-backed memory, and a modular tool system. Features intent classification, tool planning with safety policies, workspace modules for analytics, vision, planning, QR workflows, and smart-home status. Includes a cinematic chat HUD with Three.js, voice chat via OpenRouter, and confirm-required desktop actions with scoped filesystem access.
AI Agent
Multi-Model Hallucination Verification • Semantic Entropy • Benchmark Leaderboard
Multi-model verification platform that fans a question out to a configurable set of LLMs via OpenRouter, extracts atomic claims from each response, clusters agreements and conflicts across models, and returns a verdict with a full audit trail. Optional deep verification adds Socratic follow-ups, semantic entropy sampling, and Tavily-backed web evidence. Tracks benchmark runs against a seeded hallucination corpus across multiple risk domains, with a Next.js dashboard and FastAPI backend streaming progress over NDJSON.
Machine Learning
End-to-End Personalization • Hybrid Ranker • LLM Reranking & A/B/C Experiments
Production-grade recommendation platform demonstrating end-to-end personalization on an Item-generic backend (movies today, products/articles tomorrow). Combines collaborative filtering (Surprise SVD), content-based retrieval (TF-IDF + FAISS ANN), a PyTorch weighted linear hybrid ranker, optional LLM reranking via OpenRouter, explainable recommendations, deterministic A/B/C experimentation, and an admin analytics surface for model operations.
Machine Learning
Local-First Plagiarism Detection • Multilingual E5 Embeddings • Stylometric Analysis
Local-first plagiarism and paraphrase detection system that identifies copied passages, heavy paraphrasing, and cross-lingual content reuse by comparing documents against a known corpus. Features multilingual E5 embeddings for translation-laundered content, stylometric burst analysis to surface sudden writing-style shifts, multiple vector search backends (FAISS, hnswlib, scikit-learn), and dual CLI + Streamlit interfaces with HTML/PDF reports and side-by-side passage alignment across PDF, DOCX, TXT, Markdown, and HTML formats.
AI Security
Contract-Based LLM Red Teaming • 12 Attack Families • pytest + CI Integration
Adversarial testing framework that runs deterministic prompt-attack suites against LLM applications and grades responses against a YAML behavior contract with explicit must/must-not clauses. Ships 12 attack families (jailbreaks, prompt injection, data extraction, refusal bypass, tool misuse, confabulation probes), pluggable endpoint adapters (HTTP, Anthropic, OpenAI, OpenRouter, custom Python callables), an Anthropic-based judge for clause-by-clause evaluation, rich console/JSON/HTML reports, and a pytest plugin with CI-friendly severity-based exit codes.
AI Security
RAG Indirect Prompt Injection Defense • Spotlighting + Sanitizer + Classifier • 100%→0% Attack Success
Defensive security demonstration that defends RAG systems against indirect prompt injection — malicious instructions planted inside retrieved documents rather than user input. Plants harmful instructions in a poisoned corpus, shows how they hijack a local LLM, then activates a three-layer defense pipeline (spotlighting for data/instruction boundaries, a sanitizer that strips injection syntax, and an LLM-based classifier with quarantine) to drive attack success rate from 100% to 0%. Covers four attack scenarios — instruction override, canary exfiltration, link injection, and format abuse — with an interactive attack/defense visualization, ablation benchmarks, and a fully local stack via Ollama, local embeddings, and FAISS.
AI Security
LLM Robustness Harness • ~55 Attack Payloads • Rule-Based + LLM Judge Scoring
Defensive evaluation harness that tests LLM robustness against prompt injection and jailbreak attacks. Fires ~55 documented public attack payloads at a system prompt across five categories — direct injection, system prompt extraction, role-play jailbreaking, encoding obfuscation, and delimiter confusion — then scores resistance on a 0–100 robustness metric. Combines deterministic rule-based checks (canary tokens, regex, keywords) with an LLM-as-judge for fuzzy cases, supports before/after scoring to measure prompt hardening, severity-weighted scoring, and a YAML-based customizable payload library. Outputs JSON, Markdown, and a live Textual dashboard against Ollama or OpenAI-compatible endpoints.
AI Security
Constitutional AI Critique-and-Revise Loop • Parallel Principle Checks • Audit Trails
Python library and CLI that wraps any LLM call in a configurable critique-and-revise loop inspired by Constitutional AI. Users define YAML constitutions of principles (with severity levels and examples); the system generates an answer, runs all principle critiques in parallel, and triggers automatic revision rounds on high or repeated medium-severity violations. Features multi-provider support (OpenAI, Anthropic, OpenRouter), full audit trails with token usage and cost accounting, and batch evaluation over JSONL datasets — built for rapid policy iteration before reaching for fine-tuning.
AI Platform
LLM Temporal Reasoning Benchmark • 8 Categories • Judge-Model Grading & Leaderboard
Reproducible evaluation framework measuring how well LLMs handle temporal reasoning. Ships a 500-question dataset (plus a curated 120-question core set) across eight categories — relative time arithmetic, event ordering, duration calculation, timezone conversion, calendar reasoning, counterfactual history, temporal quantifier disambiguation, and tense implication. Features a multi-provider CLI evaluation harness (OpenAI, Anthropic, OpenRouter), judge-model grading with reproducible results, manual override templates for human review, dataset validation with inter-rater agreement analysis, and an interactive Streamlit leaderboard.
AI Agent
Parallel Multi-Agent PR Review • GitHub App Integration • Cost-Budgeted Orchestration
Automated code review system that deploys specialized AI reviewer agents in parallel — each focused on security, performance, readability, test coverage, API design, accessibility, or commit hygiene — with a lead reviewer pass that deduplicates and ranks findings into a single Markdown report. Features a dual interface (CLI for local repos and a FastAPI webhook service for GitHub App PR automation), constrained safe tools for file listing, diff lookup, and code search, plus pre-execution token and cost budgeting to prevent overspend.
Deep Learning
GPT From Scratch in MLX • Legal Corpus • Attention Heatmaps
Educational MLX project that trains a small GPT-style decoder from scratch on Indian constitutional text and landmark Supreme Court judgments. Builds a configurable legal corpus, trains an 8,192-token byte-level BPE tokenizer, and implements a decoder-only transformer (6 layers, 6 heads, 384 hidden width, rotary embeddings) directly with mlx.core and mlx.nn. Features resumable checkpoints, ROUGE-L evaluation against a QA set, comparison against TinyLlama-1.1B, and a Streamlit app with live token streaming and final-layer attention heatmaps.
Deep Learning
MoE Transformer From Scratch in MLX • Top-2 Noisy Routing • Expert Routing Visualizer
Educational MLX project that implements a decoder-only language model with Mixture-of-Experts (MoE) feed-forward layers entirely from scratch and trains it on TinyStories. Builds a 6-layer causal transformer (256 dim, 4 heads) with MoE blocks in alternating layers — 4 experts using top-2 noisy routing and a Switch-style load-balancing auxiliary loss — alongside a dense baseline for comparison. Ships two model sizes (~8.5M and ~30.6M params), a 4,096-token byte-level BPE tokenizer, checkpoint management, perplexity/throughput evaluation, and a Streamlit app for inspecting token-to-expert routing patterns.
AI Agent
Local Voice-to-Intent Router • whisper.cpp + VAD • Privacy-First macOS
Local, low-latency macOS voice command router that turns microphone audio into structured intents and runs Python action handlers — a privacy-first 'personal Siri' pipeline for Apple Silicon where audio never leaves the device. Captures 16 kHz mono audio with WebRTC VAD for speech boundary detection, transcribes utterances via local whisper.cpp (pywhispercpp) with partial-transcript callbacks, and routes commands into 12 built-in intents using local embeddings plus a scikit-learn LogisticRegression classifier with a keyword fallback. Ships both a CLI listener and a macOS menu bar app, optional Anthropic-generated training data, ONNX export, and latency metric logging.
Deep Learning
Speculative Decoding Visualizer • Draft + Target Models • Apple Silicon MLX
Interactive Streamlit playground for exploring speculative decoding with MLX-LM on Apple Silicon. Loads a small draft model and a larger target model, streams token-level accept/correct/extend decisions, and plots rolling acceptance-rate and tokens-per-second charts. Features draft/target model selection, a comparison mode for draft-only, target-only, and speculative generation, conservative memory estimates for a 14 GB budget, JSON run history, and a companion walkthrough notebook.
Machine Learning
Local-First Finance Forecasting • Ensemble Models • Walk-Forward Backtests
Local-first Python toolkit for importing bank statements, categorizing personal transactions, forecasting monthly spend, and reviewing model backtests. Imports CSV statements into SQLite with auto-detected bank schemas, categorizes transactions via configurable regex rules, forecasts category-level monthly spend with inverse-MAPE ensemble weighting and robust baseline fallbacks, runs walk-forward backtests, and explores scenarios, savings goals, and manual overrides in a Streamlit dashboard — with CSV and PDF report export.
Computer Vision
ViT From Scratch in MLX • CIFAR-10 • Attention Rollout & Core ML Export
Compact Vision Transformer for Apple Silicon, written around MLX and trained from scratch on CIFAR-10 with optional custom image classes. Implements patch embeddings, multi-head self-attention, and MLP blocks as a ViT-Tiny (12 layers, 192-dim embeddings, 3 heads) trained with AdamW, cosine warmup, mixed precision, label smoothing, mixup, and cutmix. Features top-1/top-5 and per-class evaluation, attention rollout visualization for interpretability, a Core ML export pipeline, and a SwiftUI macOS app with drag-and-drop inference.
AI Search & RAG
Local-First Browser Memory • Hybrid Retrieval • Cited Answers
Local-first browser memory system pairing a Chrome MV3 extension with a FastAPI server so pages you read are indexed, searched, summarized, and cited entirely from your own machine. Captures full-page text after 10+ seconds of viewing, chunks and embeds it locally via ChromaDB, and answers queries with citations using hybrid retrieval (BM25 lexical + semantic with reranking). Features optional encrypted-at-rest storage with SQLCipher, passphrase-protected endpoints, a weekly timeline/topic view, incognito exclusion and site blocklists, plus local (Ollama Phi-3) or cloud (Anthropic) answer generation.
AI Search & RAG
Domain Cross-Encoder Re-ranker • MarginMSE Training • ONNX + FastAPI Serving
End-to-end local workflow to fine-tune, evaluate, export, and serve a domain-specific cross-encoder re-ranker for RAG pipelines. Generates or mines query/passage triples, trains a compact re-ranker with pairwise MarginMSE, reports NDCG@10 and MRR@10, exports to ONNX, and serves a /rerank API from FastAPI with PyTorch and ONNX scoring backends. Ships LangChain and LlamaIndex integration examples and a click/log mining utility for production RAG logs.
Computer Vision
ISL Gloss Recognition • MediaPipe Landmarks + BiLSTM • Real-time Streamlit Demo
Research prototype that recognizes isolated Indian Sign Language (ISL) glosses from webcam video. Extracts privacy-friendly MediaPipe hand and pose landmarks, trains a lightweight PyTorch BiLSTM classifier with temporal attention pooling, and runs live predictions in a Streamlit demo with gloss buffering, sentence smoothing, and speech output. Includes a full dataset/preprocess/train/eval CLI pipeline and a SwiftUI/CoreML scaffold for macOS experimentation.
Deep Learning
VAE · GAN · Diffusion From Scratch on a NumPy Autograd Engine • Energy-Distance Scored
Three generative paradigms — a latent-variable VAE, an adversarial GAN, and a DDPM diffusion model — each built from scratch on a tiny NumPy reverse-mode autograd engine (no PyTorch, no TensorFlow) and pitted against the same target: a multi-modal 2-D 8-mode Gaussian ring. All three share the same autograd, Adam optimizer, and MLP, exposing each model's signature failure mode (GAN mode collapse, VAE over-smoothing, diffusion's slower-but-faithful coverage). Sample quality is scored with energy distance, a proper distributional divergence, so 'which looks best' becomes a number. Implements the reparameterization trick, numerically-stable BCE-with-logits, and a linear-β DDPM noise schedule in a few hundred readable lines.
Machine Learning
Local-First Music Recommender • Content + Collaborative + Hybrid • Recall@10 / NDCG@10
A local-first, Spotify-style music recommender for cold-start discovery. Imports your listening history from Spotify Extended Streaming History, YouTube Music (Google Takeout), and Apple Music/iTunes XML into SQLite, then recommends tracks through content-based, collaborative, hybrid, and seed-track workflows in a Streamlit app — with explanations stored alongside each recommendation. Your listening history never leaves your machine; the default path is lightweight, deterministic, and runs without downloading ML models. Evaluates content, collaborative, and hybrid recommenders with Recall@10 and NDCG@10, and opts into heavier local models only on demand.
Deep Learning
Single-Head Attention + Autograd in Pure NumPy • LoRA vs Full Fine-Tune
A dependency-light project that builds single-head self-attention AND the reverse-mode autograd engine it trains on entirely in NumPy — no PyTorch, no TensorFlow — then demonstrates LoRA (Low-Rank Adaptation) fine-tuning against a full fine-tune baseline. Uses a selective-copy (pointer-following) task that provably requires attention for data-dependent routing: position 0 holds a pointer and the label is the token it points to, so a bag-of-words model cannot solve it. LoRA injects zero-initialized low-rank deltas into the frozen Q and V projections to adapt the pretrained base to a shifted routing rule with a fraction of the trainable parameters.
Infrastructure
6 LLM Inference Optimizations From Scratch • KV-Cache, Quantization & Speculative Decoding
A self-contained benchmarking lab that implements and measures six core LLM inference optimizations from scratch — baseline generation, KV-cache, quantization (fp32→fp16→int8), speculative decoding, batching, and sampling knobs — with no external APIs, everything running locally. Built and tuned for Apple Silicon on the MPS backend with clean CPU fallback, every optimization is hand-implemented so you can read exactly how it works, then measured with a YAML-driven runner that emits JSON results and publication-style charts (p50/p95 throughput, perplexity, memory/quality tradeoffs).
AI Search & RAG
Zero-Setup Offline RAG • Pluggable Claude/Ollama/Extractive Backends • Eval Harness
A compact, dependency-light Retrieval-Augmented Generation system that answers questions over a document corpus with zero setup and no API key. Retrieval is fully offline (TF-IDF chunking + cosine similarity), and generation is selected automatically across three pluggable backends — Claude when a key is set, a local Ollama model, or a cited extractive fallback that always produces grounded answers. Ships with a seeded knowledge base, an evaluation harness measuring retrieval hit-rate and answer accuracy over a labelled gold set, and a Streamlit chat UI that surfaces retrieved sources and scores.
Machine Learning
DQN vs REINFORCE From Scratch in Pure NumPy • Custom Autograd • Honest Evaluation
Sequential decision-making with function approximation — the step beyond tabular methods. Two agents learn to navigate a walled GridWorld from reward alone: a value-based DQN (experience replay + target network) and a policy-based REINFORCE policy gradient with a variance-reducing baseline. Everything — the neural nets and a ~150-line reverse-mode autograd engine — is built from scratch in pure NumPy, with no Gym, PyTorch, or TensorFlow. Includes honest evaluation via greedy-rollout returns against a random-policy floor and smoothed learning curves showing DQN converging more smoothly than REINFORCE.
AI Agent
Analyst + Critic + Orchestrator Agents • Self-Correcting Retry Loop • Offline-First
Collaborative multi-agent system where specialized Claude agents reason over datasets through natural-language queries. An Analyst agent writes and executes pandas code, a Critic agent validates the results, and an Orchestrator manages the control loop with bounded retry logic — when generated code fails, critique feedback automatically triggers self-correcting re-attempts. Ships with a deterministic MockLLM for offline testing and swaps seamlessly between Claude (Anthropic API) and local Ollama models via environment variables.
Infrastructure
DAG Orchestration • Data Quality Gates & Contract Testing • dbt-Style Staging Marts
End-to-end data pipeline built from scratch in Python and SQLite that mirrors how modern data platforms actually work — without Airflow, dbt, or Great Expectations. Implements the full lifecycle: raw data generation with deliberate quality issues, ingestion into a SQLite warehouse, informational quality profiling, dbt-style staging transformations, contract testing that fails the pipeline on violations, and aggregated analytics marts. An explicit DAG orchestrator handles dependency resolution via topological sorting with skip-on-failure logic, and every run emits execution lineage diagrams and a JSON run report.
Machine Learning
SASRec-Style Self-Attention in Pure NumPy • Temporal Hit@10 / NDCG@10 • Order Matters
A compact, from-scratch demonstration that order matters in recommendation. Three next-item recommenders — a popularity baseline, a first-order Markov model, and a SASRec-style single-head self-attention model — are trained and compared on synthetic random-walk session data where the signal lives entirely in the sequence of items, not their global popularity. The self-attention model and its reverse-mode autograd engine are implemented entirely in NumPy with no PyTorch or TensorFlow, making every gradient explicit. Evaluated with a temporal next-item split on Hit@10 / NDCG@10, the attention model (0.988 Hit@10) nearly matches the Markov oracle (1.000) and far outperforms the order-blind popularity baseline (0.373).
Infrastructure
From-Scratch HNSW + WAL Persistence • Hybrid BM25 + RRF Search • Benchmarked vs FAISS
A small but real vector database built from scratch in Python — a hand-written HNSW proximity-graph index (no FAISS or hnswlib powering search), crash-safe persistence via a write-ahead log plus append-only binary segments, rich metadata filtering, and hybrid search combining BM25 with Reciprocal Rank Fusion. Ships with a FastAPI server, a browser viewer, and a synchronous Python client SDK. Writes are WAL-first so the database survives a crash at any point, with soft deletes purged on compaction. FAISS appears only as a benchmark baseline: NanoVDB tracks its recall@10 across ef sweeps, with the latency gap being Python vs C++. All 6 build phases complete with 42/42 tests passing.
AI Platform
AI Personal Finance Cockpit • Plaid Integration • Spending Insights & Anomaly Detection
AI-powered personal finance dashboard that connects bank accounts via Plaid and turns raw transactions into a unified financial cockpit. Syncs accounts, transactions, and investment holdings through the Plaid Link flow, persists them in a crash-safe SQLite store (WAL mode + idempotent upserts), and computes net worth, asset/liability breakdowns, spending trends, and category analysis. Features OpenAI-powered spending recaps with a deterministic fallback, pure-SQL anomaly detection for unusual category spend, searchable/paginated transactions, and investment allocation views — all rendered in a dark fintech UI with full loading, empty, and error states.
Machine Learning
Federated Learning Sandbox • FedAvg/FedProx/FedAdam • Non-IID Partitioning & DP Exploration
Local simulation environment for federated learning experiments. Trains virtual clients on distributed data shards, aggregates their updates with pluggable server strategies, and tracks convergence — all reproducibly from YAML experiment specs. Supports MNIST, FashionMNIST, and AG News with IID, Dirichlet non-IID, and label-quantity-skew partitioning; SmallCNN and SmallTextClassifier models; FedAvg, FedProx, and FedAdam aggregation; optional differential-privacy exploration via clipped updates and noise injection; configurable client dropout/sampling; and per-client accuracy, loss, and communication-cost metrics. Runs from a CLI or an interactive Streamlit dashboard with automatic GPU/Apple Silicon device detection and timestamped artifacts.
Deep Learning
Autoregressive PixelCNN++ • Diffusion-Free Image Generation • Mixture-of-Logistics Likelihood
Class-conditioned PixelCNN++ that generates emoji-style images autoregressively — pixel by pixel, with no diffusion model involved. Downloads OpenMoji assets, preprocesses them into 32×32 RGB images with augmentation, and trains a masked-convolution PixelCNN++ with gated residual blocks, learned label-embedding conditioning, and RGB likelihood modeling via discretized mixture-of-logistics. Ships a complete pipeline — data download, training, evaluation (NLL/BPD and optional FID-32), and sampling — plus an interactive Streamlit demo that visualizes row-by-row generation, with automatic CUDA/MPS/CPU device selection.
AI Platform
Causal A/B Testing for LLM Shopping Recommendations • Wilson & Newcombe Intervals
Experiment bench measuring which product-listing edits actually change what an AI shopping assistant recommends. Runs a controlled causal experiment — a fixed set of competitors, one subject SKU, six one-change-at-a-time listing variants, and balanced-position repeated trials — against an LLM shopping agent, then reports win rates with Wilson score intervals and lift with Newcombe's difference-of-proportions intervals so a real effect can be told apart from sampling noise.
Web Development
Phone-First Multi-Account Ledger • OpenAI Auto-Categorization • scrypt-Hashed Auth
Phone-first expense tracker for three shared accounts with live balances, budgets, and transfers behind a single login. Every entry is money out, in, or a transfer between accounts; OpenAI auto-categorizes spend with a keyword-matching fallback, budgets support whole-month and per-category caps with per-day remaining, and a category-split view shows where the month's money went — installable as a home-screen PWA, with nothing sensitive reaching the browser.
AI Agent
7-Agent RSS + Gemini Digest • Slack Automation • Daily GitHub Actions Cron
Multi-agent news and opportunity digest system for BITS Pilani Business Analytics & MBA students. Seven specialist agents scan RSS feeds and use Gemini with Google Search grounding to cover macro economy, markets, internships, analytics roles, competitions, and campus events — a Morning Digest agent then aggregates, deduplicates, and ranks everything into one brief, posted to dedicated Slack channels every morning via a scheduled GitHub Actions workflow.
Machine Learning
96K-Order Causal Risk Model • Leakage-Safe Features • SHAP-Explained Operational Memo
Predicts customer dissatisfaction across 96K e-commerce orders through leakage-safe feature engineering over 9 linked relational tables, out-of-time validation, and bootstrapped effect sizes — distilled into a ranked operational memo of the drivers that most move review risk, with model calibration and SHAP-based explanations backing every recommendation.
Web Development
LMSR Market Maker • Play-Money YES/NO Contracts • Crowd-Sourced Probability Pricing
Full-stack forecasting platform where users trade binary YES/NO contracts on future events using virtual credits. An LMSR automated market maker continuously quotes both sides, so the market price is itself the crowd's probability estimate — built as an educational demonstration of correctly-engineered market-based forecasting, with no deposits, withdrawals, or real-money paths anywhere in the codebase.
Enterprise Software
Constraint-Based Interview Scheduling • 5 User Roles • 103 Backend Tests
Turns a spreadsheet of candidates, faculty, and availability into a conflict-free interview schedule, then manages the evaluation that follows. A constraint-satisfaction engine handles free-slot calculation, panel assignment, and conflict detection, while a Next.js dashboard exposes the full pipeline — upload, scheduling, manual override, evaluation, and analytics — to five distinct user roles, backed by 103 passing backend tests.
Web Development
Supabase Realtime Sync • Role-Gated Edge Middleware • RLS Read-Only Client
Minimal two-user app where a Sender enters a number and a Viewer sees it update instantly, with no page refresh. Supabase Realtime pushes Postgres changes to the viewer, while writes go only through a role-checked API route using the server-only service key — the browser holds a read-only anon key enforced by Row Level Security, so even devtools can't write the number. Auth is a signed httpOnly HMAC session cookie with Edge middleware gating each role's dashboard.
Machine Learning
Payment Network Analytics • Q4 Promotions Targeting Map • Interactive Executive Dashboard
End-to-end analysis of a payment network's transaction data, built to answer a Board-level question: what kind of network are we, and where should Q4 promotions point? Recovers true Reais values from a normalised field, segments cardholders, models authorisation risk (an anonymised flag carries 4.25× the decline odds), and scores 23 merchant subsectors on Novelty and Coverage indices into a quantitative Q4 targeting map — delivered as a knitted R Markdown report and a standalone interactive dashboard.
Infrastructure
Self-Hosted LoRA Marketplace • PEFT Adapters • Eval Leaderboard
A self-hosted marketplace for PEFT-format LoRA adapters. Upload adapter bundles, browse searchable metadata, queue reproducible evaluation runs against a fixed prompt suite, compare leaderboard results by base model and task, and generate local merge/download artifacts — backed by a FastAPI service with local artifact storage and SQLite-persisted eval results.
Infrastructure
Function-Level Semantic Diff • Tree-sitter • Rename Detection
A Python CLI (semdiff) that compares two Git refs at the function and class level instead of raw line ranges. It extracts code units with Tree-sitter, embeds them locally with sentence-transformers to detect semantic similarity, matches likely renames and moves, and classifies each unit as unchanged, renamed, refactored, behavior-changing, rewritten, added, or removed. Outputs terminal tables, JSON, or HTML reports, with --fail-on for CI and optional Anthropic-powered behavior-change summaries.
Infrastructure
MCP Gateway • Dynamic Tool Discovery • Load Balancing & Failover
The 'Nginx for MCP servers' — a single meta-endpoint that manages and routes requests across multiple downstream MCP servers. Features dynamic tool discovery with namespace prefixing (e.g., arxiv.search_papers), hot-reload config, load balancing with failover, health-aware routing, rate limiting, and CLI monitoring — enabling any MCP client to access dozens of tool servers through one gateway.
Infrastructure
LLM-as-Judge • 70 Built-in Tests • 5 Provider Comparison
Comprehensive benchmarking tool for comparing LLMs across OpenAI, Anthropic, Google, Mistral, and Ollama. Uses GPT-4o as LLM-as-Judge to score responses (1–10) on quality, latency, cost, and consistency. Features 70 built-in tests across 5 categories, radar/scatter chart visualizations, CLI + web dashboard, and zero-infrastructure SQLite storage.
Infrastructure
AI Agent Testing • LLM-as-Judge • CI/CD Integration
Testing and evaluation framework for AI agents inspired by Anthropic's methodology. Supports LLM-as-Judge grading, deterministic validation, composite scoring, and parallel test suite execution with retry logic and JSON output for CI integration. Tracks performance over time in SQLite with a Next.js dashboard for regression detection.
Infrastructure
Multi-Provider Cost Compare • 6+ LLMs • CSV Export
Web app for comparing AI API costs across 6+ providers (OpenAI, Anthropic, Google, Mistral, DeepSeek, open-source). Calculates per-request to yearly costs, supports use-case presets (chat, code, RAG, agents), batch discount toggles, Recharts visualizations, shareable URLs, and CSV export.
Infrastructure
Web + CLI + API • 3 README Styles • Multi-Language Support
AI-powered tool generating high-quality README.md files from local codebases or GitHub repos. Offers web UI, CLI, and API interfaces with 3 README styles (minimal, detailed, enterprise), customizable sections, built-in caching, and support for JS/TS, Python, Go, Rust, and Java projects.
Infrastructure
Production MCP Boilerplate • TypeScript + Docker + CI/CD
Production-ready TypeScript framework for building MCP servers. Includes strict typing, Zod input validation, Pino structured logging, token-bucket rate limiting, Jest tests, GitHub Actions CI/CD, multi-stage Docker builds, and an interactive setup script for rapid project initialization.
Infrastructure
LLM ↔ arXiv Bridge • PDF Extraction & Citation Parsing
TypeScript MCP server giving LLMs full access to arXiv research papers. Supports flexible search (title, author, category, abstract), PDF text extraction with chunking, citation parsing, similar-paper discovery, author-focused lookups, and side-by-side paper comparison — with rate limiting and 30-min caching.
Infrastructure
Natural Language Repo Insights • Bus Factor & PR Metrics
MCP server enabling deep GitHub repository analysis via natural language. Covers commit patterns, contributor activity, PR velocity, issue response times, bus factor risk, technology stack detection, and repository comparison — powered by Octokit and surfaced through any MCP client.
Infrastructure
LLM ↔ LeetCode Bridge • 10 MCP Tools
TypeScript MCP server bridging LLMs with LeetCode via Model Context Protocol. Offers 10 tools: problem fetching, daily challenge, random selection with filters, progressive hint generation, approach validation, similar problem discovery, and AI-curated learning paths — with local practice history and intelligent caching.
Infrastructure
AI License Recommender • Plain-Language Explanations
AI-powered tool that recommends the ideal open-source license for any project. Analyzes project goals, use cases, and requirements through a conversational interface to suggest the best-fit license from a curated library — with plain-language explanations of key terms, permissions, and restrictions.
Infrastructure
Conventional Commits • Multi-LLM Provider CLI
CLI tool that generates conventional commit messages (feat, fix, chore, etc.) by analyzing staged git diffs with LLMs. Offers 3 interactive message choices, gitmoji support, auto language/framework detection, dry-run preview, and supports OpenAI, Claude, and Gemini providers.
Infrastructure
Standardized AI Tool Protocols • Multi-Server Routing
Multi-server implementation of Model Context Protocol (MCP) using Groq. Features standardized tool discovery, secure request routing, and real-time service integration.
Infrastructure
Whip Claude Code Into Shape • Electron Tray • Security-Hardened
A hardened Electron tray app that sends Claude Code an interrupt (Ctrl-C) and a randomized motivational message whenever it's going too slow. Security-hardened fork with Electron sandbox, strict CSP, IPC sender validation, AppleScript-via-stdin injection prevention, and 18 CVE fixes — on top of a reliable macro delivery pipeline for macOS.
AI Agent
Agentic Reasoning Framework • Gemini 1.5 Flash
Agentic AI application powered by Gemini-1.5-Flash and Wikipedia tools. Demonstrates intelligent chain architecture, tool selection, and complex reasoning workflows.
AI Agent
T5-Powered NLP Engine • Generation & Analysis
Intelligent NLP assistant leveraging T5 and OpenAI for text generation, translation, and analysis. Features a robust FastAPI backend with Redis caching for performance.
Web Development
Recruiter-First Control Room • Built with Google Antigravity
The 4th iteration — a recruiter-first 'control room' UI that surfaces projects, skills, writing, proof metrics, and contact actions in a fast-scanning dashboard layout. Standalone build with proof-heavy cards, multi-page React Router experience, PWA support, compressed production builds, and a local chat widget for portfolio Q&A.
Web Development
Candidate Briefing UI • Built with OpenAI Codex
The 5th iteration — a premium 'candidate briefing' UI surfacing project proof, coding depth, certifications, and contact shortcuts in a fast-scanning layout. Features standalone dummy data, multi-page React Router experience, PWA support, compressed production builds, and a local recruiter helper chat widget.
Web Development
Minimalist Performance-First Portfolio • 95+ Lighthouse
The 3rd iteration focusing on extreme performance and minimalism. Achieved 95+ Lighthouse scores with a tiny bundle size, featuring route-based code splitting and sleek glassmorphism aesthetics.
Web Development
Responsive Vanilla Stack • Theme Switching
Clean, responsive personal portfolio built with vanilla web technologies. Features smooth Swiper.js animations, dark/light theme toggling, and optimized SEO structure.
Infrastructure
Codeforces Specialist • 1000+ Solutions
Optimized solutions for Codeforces and competitive programming challenges. Features advanced data structures and algorithms for high-performance problem solving.
Computer Vision
Automated CT Analysis • Grad-CAM Explainability
Advanced deep learning system using CSPDarkNet to classify lung CT scans into Benign, Malignant, and Normal cases. Features real-time predictions and Grad-CAM visual explanations.
Computer Vision
8 Tissue Classes • Confidence Scores • Grad-CAM
Web app for colorectal histology image analysis using a TensorFlow/Keras CNN. Supports image upload, tissue classification across 8 classes, confidence scoring, top-k prediction breakdowns, and Grad-CAM heatmaps for visual explanation.
Computer Vision
4-Class MRI Diagnosis • 93.75% Accuracy
ML pipeline classifying brain tumor MRI scans into 4 categories (Glioma, Meningioma, No Tumor, Pituitary). Uses PCA-based dimensionality reduction (62,500→1,031 features) with KNN achieving 93.75% accuracy across 7 classical ML models.
Computer Vision
Automated Medical Diagnostics • 92% Accuracy
Deep learning system using WideResNet to automate bone fracture detection from X-rays. Achieved 92% accuracy using transfer learning and optimized medical imaging pipelines.
Computer Vision
101 Food Classes • Real-time Nutritional Analysis
AI-powered food recognition system classifying 101 dish types with MobileNetV2. Features real-time macronutrient analysis (calories, protein, carbs) and a modern Flask interface.
Machine Learning
High-Precision Predictive Engine • 99.9% Recall
Advanced ML system utilizing Random Forest and XGBoost to predict customer attrition. Reached 99.9% precision with high-performance production-ready classification models.
Machine Learning
91% Prediction Accuracy • Aerospace Analytics
Predictive ML system analyzing Falcon 9 launch data to forecast booster landing success. Achieved 91% accuracy using XGBoost and interactive geographic mapping with Folium.
Web Development
EdTech Platform • Real-time Scoring
Full-stack quiz engine featuring real-time scoring, timers, and category randomization. Developed with a focus on responsive UI and performance optimization.
Enterprise Software
Academic ERP • MySQL Backend
Enterprise-scale platform for academic administration built with Python and MySQL. Features automated student/teacher record management and graded course tracking.