Projects

AI agents now run commands, edit files and talk to production systems on their own. When something goes wrong, the agent's own log is the only record, and the agent can edit it. I built a recorder the agent can't switch off or rewrite.

• Every agent action (session start, prompt, each tool call and its result, stop) is captured by a system-level hook the agent has no permission to change or disable.
• Each event is hash-chained and signed with a key the agent never sees, then written to an append-only journal on a separate machine. Not even the collector, or root, can rewrite it.
• No secrets or personal data leave the agent's machine: only metadata, hashes and redacted command text.
• A verifier replays the chain, checks every signature, and cross-checks the agent's own transcript. Delete a single tool call from the agent's log and it reports exactly what's missing.
• Hardened against the agent itself: it ignores the agent's environment variables, proxy settings and Python path tricks, which are the ways an agent could quietly blind its own recorder.

Live in a multi-agent fleet, piloted on one production agent first. From idea to a verified pilot in one day.

A personal AI agent that lives on your own computer and reaches your phone from anywhere. No router to configure, no account to make, nothing stored on anyone else's servers.

• Pair your phone by scanning a code. The phone side is a web page, with no app store.
• End-to-end encrypted: messages cross an encrypted relay that can't read them.
• Remembers you, in plain markdown files on your own computer that you can open and edit.
• Send it photos and files from your phone.
• Works with your Claude account, or with any OpenAI-compatible endpoint, even a model running on your own machine. Mine runs Gemma 4 26B, a mixture-of-experts model, at 8-bit, on a Mac mini M4 Pro; the model itself never touches the internet.
• Free installers for Mac (Apple silicon and Intel), Windows and Linux.

Hook-based out-of-process context engineering middleware for Claude Code. The model doesn't know it's been augmented — the hook runs first, curates what context reaches the prompt, and reasoning stays cleanly separated from curation. That one design choice unlocks three things at once: universal applicability across every model, skill, and subagent in a Claude Code session; clean replaceability on either side of the boundary; and auditability at the trust boundary — the right shape for CUI-handling, CMMC, and FedRAMP work.

Built in Rust. Persistent memory, checkpoints, and lessons via amp-rs. Semantic code search. A knowledge graph that strengthens connections over time. The Stanford IRIS / MIT Meta-Harness paper named the category last month; Nellie is my implementation of it. v0.5.3 installs in two lines.

The Rust reference implementation of Agent Memory Protocol — structured, persistent memory through three primitives: checkpoints, lessons, and memory. Local-first, fast, and built to be embedded anywhere. Powers Nellie's memory layer.

The open spec behind amp-rs. Three primitives — checkpoints, lessons, and memory — that let AI agents pick up where they left off instead of starting from zero every session.

Open-source corporate AI gateway with full audit trail on every LLM interaction. Five layers: Traefik for traffic management, Keycloak for identity and SSO, LiteLLM for unified model routing, Langfuse for observability, and content safety guardrails. Users set one environment variable and their workflow doesn't change — but every query is logged, every cost is tracked, every policy violation is flagged.

Supports both API key and subscription passthrough (Claude Max, Teams). Replaces commercial gateways that charge $2K-20K/month. Runs on a laptop.

AI coding assistants lose context between sessions, skip steps, and produce inconsistent code. DevPlan fixes this with structured planning, automatic validation, and a lessons-learned system that prevents the same mistakes from happening twice.

21 tools across planning, generation, validation, and issue remediation. It interviews you about your project, generates a validated plan, then hands off to AI for mechanical execution — while a separate verifier agent actively tries to break the result.

Works with Claude Code, Cursor, Aider, Cline, and Windsurf. Open source, MIT licensed, runs on Cloudflare Workers.

The boredom engine. A proactive agent daemon that observes, orients, decides, and acts — instead of waiting for input. OODA-loop architecture for agents that take initiative.

Memory Exchange & Sharing Hub — a secure federation protocol for AMP nodes. Share agent memory across teams without giving up control of your data.

Radarr/Sonarr for LLM models. Monitors HuggingFace for new releases matching a watchlist and auto-downloads them to a local library. Set it and forget it.

Drop-in Microsoft Graph API mock server for testing. Returns realistic JSON fixtures for 50+ endpoints across 3 cloud targets (Commercial, GCC, GCC-High) and 3 scenarios (greenfield, hardened, partial deployment). Point your Graph client at localhost instead of graph.microsoft.com — same response shapes, zero auth required.

Supports stateful mode for deploy-then-verify workflows, $filter/$expand/$top query parameters, and error simulation (throttling, 403s, 500s). Built for testing CMMC compliance tooling without touching a real tenant.