Full-stack RAG pipeline that clones GitHub repos, AST-chunks code via tree-sitter across 13 languages, and answers questions with cited, traceable sources using hybrid vector + keyword search (RRF fusion) and cross-encoder reranking. Stress-tested against 12 real public repos (297K+ chunks, 94K+ files, 4.3M+ lines of code) with per-repo PostgreSQL list-partitioning and 65% Recall@8 on a hand-labeled eval set. Indexing runs as a resumable BullMQ worker that skips unchanged repos via commit SHA and resumes per-file after a crash.

Turns a ChatGPT account into an OpenAI-compatible sk- API key, with an OpenAI-compatible /v1 gateway and playground. Resolves keys against a shared Cloudflare Worker (Hono) BFF, decrypts and refreshes ChatGPT OAuth tokens (AES-GCM at rest), and translates requests to ChatGPT's Responses API with streaming. 4,800+ keys generated across 2,100+ users.

Full-stack RAG platform letting users ingest documents or crawl URLs and deploy configurable AI chat agents grounded in their own knowledge base. Header-aware chunking pipeline feeding a Cloudflare Vectorize retrieval layer, behind a multi-provider LLM chain (OpenAI, Claude, Gemini, DeepSeek, Groq) with automatic failover and streamed responses.

Monitoring platform architected across Azure Functions and Cloudflare Workers with Redis-backed incident detection: 5-failure threshold, 24-hour lookback, automatic create/resolve.

ChatGPT-style branching-conversation app backed by your own ChatGPT account. Fork and merge conversations into a connected graph, with real-time multi-tab sync and streamed responses from the OpenAI Responses API. Backed by the same Cloudflare Worker (Hono) BFF for durable conversation snapshots. 3,600+ conversations created by 1,900+ users.

Distributed rate limiter surviving Redis outages: token bucket and sliding window as atomic Redis Lua scripts shared across 4 app instances behind nginx, a circuit breaker that falls back to a local in-memory limit when Redis is unreachable, and a reconciliation pass that replays outage usage back into Redis without double-counting.