Things I have built

A self-built AI pull-request review agent in the spirit of CodeRabbit: a zero-dependency review engine that reads any diff, posts inline comments plus a summary, and runs as a GitHub Action or Cloudflare Worker with your own LLM key — no server required. Named for the jeweler's lens: close, careful inspection of every diff.

Point it at any model — API, aggregator, or local — and it runs a 5-pillar battery (capability, reliability, safety, agency, economics), recording every run, task, turn, span, and token. Compare models side-by-side in a full dashboard with CLI/TUI runners, three.js 3D data views, and portable run bundles that pool into community averages.

A hybrid agentic RAG system: it reads PDFs, Office docs, markdown, code, and web pages, builds four kinds of search indexes — BM25 vectorless, LanceDB vector, an LLM-extracted knowledge graph, and read-only text-to-SQL — and an agent decides which to chain for each question, always answering with citations. Local-first, MIT.

A cache-aware LLM router with learned quality routing: one decision engine, two runtimes (pip + npm), zero dependencies, and an HTTP /decide sidecar that slots into LiteLLM and Bifrost to pick the best model per request on cost, latency, and quality.

Agents write code, AISDLC runs the process: harness-agnostic Agent Skills plus a zero-dependency CLI for EARS specs, evidence gates G1–G6, and hash-bound approvals.

Describe a product and seven specialist agents — Orchestrator, Architect, Database, Backend, Frontend, QA, DevOps — plan, build, verify, and ship a Next.js + PostgreSQL codebase, with human approval on the risky steps.