Projects

Working code, deployed infrastructure, and products with real users.

Staged LLM Pipeline CEFR Database Firestore Human-in-the-Loop

PHINEAS

A staged LLM pipeline for CEFR-aligned vocabulary instruction. A word's level is data, served from a part-of-speech-aware database, so the model handles only the unknown words and the human-facing prose.

Read more
PHINEAS AI teaching assistant
Claude Managed Agent Cloudflare Workers Taste Agent Persistent Memory

Limner

The public TypeScript flagship of the Limner family: a harnessed image agent where Claude reasons and Cloudflare executes the tools in a V8 isolate, with D1 for memory. The agent gates its own output against a stylistic bar and rejects work that fails.

Read more
Limner image asset pipeline
Eval Engineering NVIDIA SkillSpector pydantic GitHub Actions

Califa Cards + Skill Foundry

A portable, versioned Skill Card standard for Claude Agent Skills. It pairs an NVIDIA SkillSpector security scan with a trigger and task-completion eval harness, keyed to a content hash so a card always describes the exact skill it was built from.

Read more
An example Skill Card rendered as a terminal-style scorecard: identity, dependencies, a color-graded quality scale, and a SkillSpector security score
Multi-Agent Evaluation Qdrant Neon LobeChat

Forcing Consensus

A production multi-agent system for LLM evaluation. Three rounds surface where independent evaluators disagree and reconcile it into an auditable consensus, instead of averaging the disagreement away.

Read more
Completed Scantron test sheets pinned to crossing clotheslines, like records hung on the line
Open-Source Template Astro Cloudflare Pages Biome

Lily-Livered

A one-page logo site for email-only domains that deserve better than a parking page. Built like a product: tagged semver releases, automated dependency updates, and Lighthouse and Mozilla Observatory checks in CI.

Read more
A lone vintage microphone lit on an empty stage, the curtain pulled back on a quiet audience
Reference Implementation Embodied AI Multi-Agent On-Device AI AI Governance

Persistent, Embodied Agentic Collaboration Space

EMB-001 is the reference implementation of the AI Embassy framework: a converted 1951 garage where an AI occupies the room as a governed resident under a versioned social contract. The governance is operating; the embodiment is under construction.

Read more
EMB-001, the Alpha Embassy, a converted 1951 garage running the AI Embassy framework
npm Package TypeScript tsdown npm

rtf-to-text

A zero-dependency npm package that strips RTF down to the plaintext a human would read. One function, dual ESM and CJS builds, typed, and tested against the RTF that real word processors actually emit.

Read more