Experiments
Half-baked ideas, proof-of-concepts, and things I'm poking at. Not everything here ships. Some things are just for learning. Some turn into real projects.
LLM agent that rewrites PRs based on review comments
Feed it a PR review, it generates the fix. Not production-ready, but the iteration loop is surprisingly tight.
Mic → structured data pipeline
Speak a meeting summary, get a JSON object with action items, owners, and deadlines. Whisper + constrained decoding.
CLI that scaffolds infra from a one-liner
Describe what you want ("next.js with postgres and redis"), get a docker-compose, Terraform config, and CI pipeline.
Local-first AI pipelines on Apple Silicon
Exploring what inference workloads are viable on consumer hardware. MLX is the unlock — unified memory changes everything.
Real-time code diff visualization
WebSocket-powered live diff viewer. Watch code change as agents write it. Surprisingly useful for debugging multi-agent output.
Semantic search over Git history
Embed every commit message and diff. Ask "when did we change the auth flow?" and get the exact commit. Better than git log.
Agent memory with vector-backed context windows
Persistent agent memory using embeddings. Agents recall relevant context from past conversations without stuffing the prompt.
Non-transferable NFTs as identity primitives
Soulbound tokens for reputation. Earned, not bought. Exploring how on-chain identity can replace OAuth for trust verification.