I build infrastructure for AI agents, and the large-scale systems behind Amazon’s supply chain.

An LLM cache that knows when its answers went stale, and replays the reasoning instead of re-deriving it.

YORO (You Only Reason Once) is an OpenAI-compatible caching proxy. A semantic cache serves an answer whenever a request looks familiar, which means it keeps serving a dead answer once the world changes. YORO fingerprints what each answer depends on, invalidates when those inputs move, then replays the stored derivation against the new inputs. It drops in front of vLLM, llama.cpp, or any OpenAI-compatible endpoint.

Benchmarked on gpt-oss-120B and Qwen2.5-32B: accuracy under drift held at 0.96 versus 0.16 for a naive cache, at 10 to 21% of the no-cache output tokens.

An always-warm code index per repo, served to AI coding agents over MCP.

Every new coding-agent session used to rediscover the same repo from scratch. Skep runs a single-binary daemon per repo that keeps a tree-sitter and SQLite FTS5 index warm, and answers symbol search, file context, and call-graph queries over MCP without a single LLM call. On top of the index sit a task queue with dedup and classification, approval gates for large changes, and cross-repo delegation so one repo’s session can hand work to another. Local-first, no cloud, no API keys.

A local-first knowledge vault that gives AI coding agents memory across sessions.

Notes are plain HTML files in a folder you own. Weft indexes them and surfaces what your brain would recall right now, ranked by backlinks, co-access, recency, and local embedding similarity, and hands that same context to coding agents over MCP. Sync is end-to-end encrypted to a bucket you control. It ships as a macOS app, an iOS app, and a CLI.

It tracks its own builder. My project notes, decisions, and drafts live in a Weft vault, and Claude Code reads them at the start of every session. This page was written from those notes.

A social feed where every post is a runnable app.

Posts are declarative specs rendered on-device inside a sandboxed iframe, so a feed item can be a calculator, a live chart, or a small game instead of a screenshot of one. Behind it: a Go backend with event workers, a Next.js web client, native iOS and Android apps, Swift and Kotlin SDKs, on-device feed ranking, and a TypeScript MCP server so AI tools can publish straight into the feed. Live on the web, on iOS through TestFlight, and on Android.