Work

A self-hosted platform where AI agents do real work, under human authority.

My self-hosted platform for running LLM coding agents. A deterministic kernel (TypeScript + Postgres) owns the work state, budgets, and an audited event trail, then hands execution to agents, automated research, and home services. Everything ships through branch-protected GitOps on a home k3s cluster, and anything irreversible waits on my approval. The agents do the judgment work; a human keeps final authority. Built as small, permission-bounded repos instead of one monolith.

Deterministic kernel · GitOps on home k3s · ~1,300 agent-assisted PRs · human-gated

• TypeScript
• Postgres
• Kubernetes (k3s)
• Flux
• GitHub Actions
• LangGraph
• Claude Code

One audited choke point for everything an agent does to the outside world.

A policy-gated side-effect broker. Every agent-initiated external write (email, calendar, tasks, a browser action) passes through one audited gateway with an approval step. It's how agents get to act in the real world while a human still signs off on what matters.

Dual-approval gate on all outbound side effects

• TypeScript
• Policy engine
• OAuth integrations

The Kubernetes operator pattern, applied to my house.

A declarative home-automation controller that replaces brittle if-this-then-that rules with a reconciliation loop: behaviors are pure functions of a world snapshot, rendered to a desired state and diffed onto Home Assistant. Includes a weighted-evidence occupancy engine that fuses motion, BLE, and camera signals into per-room presence. Heavily typed, well-tested, and running my actual house.

Pure-function behaviors · reconciled desired state · weighted-evidence occupancy

• Python
• Home Assistant
• AppDaemon
• Kubernetes

A spoiler-aware knowledge graph over a ten-book, ~3.3M-word series.

An NLP corpus platform that extracts characters, relationships, and events from a ten-book series and serves them to LLM clients. BookNLP and transformer embeddings feed a Postgres knowledge graph, exposed through an MCP server with a suite of query tools. It runs on a deterministic-bulk-extraction plus LLM-judge-the-hard-residual architecture, with trust-tiered, era-scoped canonization.

~3.3M words · MCP server · trust-tiered canonization

• Python
• BookNLP
• Transformers
• Postgres
• MCP

A health app that became a data source for agents acting on my behalf.

A cross-platform mobile app for journaling, habits, and HealthKit data, backed by a Rails 8 API with broad Maestro end-to-end coverage. It now doubles as a structured data source for AI agents that draft entries and take action on my behalf. It is the personal-telemetry layer that grounds the rest of the system.

Expo + Rails 8 · HealthKit · Maestro E2E

• React Native
• Expo
• TypeScript
• Ruby on Rails
• Maestro

Brazilian jiu-jitsu tracking, with a nudge to actually journal.

A Rails web app for journaling jiu-jitsu training and tracking progress, with scheduled Twilio SMS reminders that prompt me to capture notes right after each session, while the details are still fresh.

Scheduled Twilio nudges to capture notes post-session

• Ruby on Rails
• Postgres
• Twilio