Code knowledge graphs for AI agents PUBLISHED

Python Static analysis Graph modeling Apache-2.0

Multi-language static-analysis engine that turns a codebase into a queryable knowledge graph — with a per-edge confidence taxonomy (EXTRACTED / INFERRED / AMBIGUOUS), git-history archaeology modelled as graph data, and automated architecture docs across six output surfaces including Obsidian vaults. Built so coding agents can reason about unfamiliar code instead of guessing at it. v0.8, shipped across an ADR-driven roadmap with 90+ logged decisions.

Plain-English E2E tests, run by an AI agent PUBLISHED

TypeScript LLM agents Playwright/CDP GitHub Actions Zod

Write the test in English; an agent runs it against a live preview of every pull request. It reads the page's accessibility tree, caches what works so steady-state runs cost zero LLM calls, and re-resolves only the step that drifted when the UI changes. Fork-safe two-workflow model lets a fork's PR post a comment and status check without ever touching repo secrets. Ships with a flake-vs-regression classifier and watchable failure replays.

Tanglish voice service-desk agent

Python Agentic workflows BM25 RAG MCP ASR/TTS

Provider-agnostic voice agent with typed ports and a zero-key local mode. BM25 retrieval, working plus long-term memory, guardrails, MCP-shaped tools, and a draft → confirm → execute flow that stops the agent making state changes nobody approved.

Streaming VAD & ASR front end, from scratch

Python NumPy SciPy scikit-learn Whisper/Saaras

Voice-activity detection built from the DSP up — hand-written features, adaptive gating, smoothing, and a learned logistic-regression path — plus a mic/file-to-ASR pipeline with Whisper and Saaras adapters. 31 passing tests.

Autonomous options trading agent HACKATHON

Python Alpaca API Options

Deadline-aware options agent that optimises realised cash rather than paper P&L. A deterministic engine makes every trade decision; the LLM is confined to risk-scaling and journaling. Built for the lablab.ai × Alpaca hackathon.

Auditable AI claims adjudication HACKATHON

FastAPI Gemma vLLM/ROCm AMD MI300X

APPROVE / FLAG / DENY decisions grounded in verbatim policy clauses, with every citation programmatically verified against the source before display — failed citations are visibly rejected, never silently dropped. One structured-output call, switchable between Fireworks AI and a self-hosted MI300X via one env var.

Pixel-art companion that runs Claude Code on Windows

Electron TypeScript React Vite

Windows port of @itmesneha's macOS agentrocky. Rocky walks along your taskbar, opens a retro terminal chat when clicked, and runs Claude Code as a persistent subprocess. The Windows-specific work was in the corners: claude.cmd resolution, transparent overlay quirks, stream-JSON CRLF tolerance, DPI-aware sprite rendering.

Multi-tenant clinic management SaaS

Next.js TypeScript Supabase Tailwind

End-to-end practice management — patient records, appointments, billing, multi-clinic tenancy. Built with Claude Code as the engineering pair. The story site walks through the build approach and the design decisions.

Real-time multiplayer chess — with text-based moves

React TypeScript Supabase WebSockets

Move pieces by typing e2e4. Real-time sync, move validation, time-limited hints, persistent profiles. Built because I wanted chess without dragging pieces.

P2P messaging with zero internet

Flutter Dart Bluetooth WiFi Direct

Android app for peer-to-peer messaging over Bluetooth and WiFi Direct. E2E encrypted, auto device discovery (100m range), voice notes, images, offline message queuing.