I build infrastructure that makes shipping software feel easy.

Authored and published an MCP server exposing five read-only, annotated tools over both stdio and JSON-RPC HTTP, with bearer-token and OIDC authentication, stdlib-only JWT verification, and a refusal to bind past loopback unauthenticated.

Listed on the MCP Registry and added to Claude Code, Claude Desktop or Cursor with a single command.

repo2graph

Built enforced token ceilings into retrieval — budgets re-measured after rendering rather than trusted, capped neighbour rows, bounded fan-out — and a skill set plus PreToolUse hook that strips noisy tool output before it ever enters a model's context.

Attacks the re-sent context pile rather than reply length, which is where the overwhelming majority of tokens actually live.

tokenmiser

Designed a retrieval layer combining BM25 lexical scoring, optional dense MiniLM fusion, and a one-hop walk across a tree-sitter code graph, chunked so every block arrives with its callers and callees in the header.

Every retrieved block is stamped with its file and line range and the edge that pulled it in, so a model's claim can be checked rather than trusted.

repo2graph

Built a seven-phase build loop — plan, test, implement, review, verify, remember, improve — where each phase is its own bounded subagent with a written contract, communicating only through a single shared state file.

Hard gates between phases: a run cannot skip its own tests, and four iterations without a clean verify hands back to the operator.

Agentic App Loop

Routed subtasks across Groq, Gemini, Mistral, Cerebras and OpenRouter by task type from a single config file, executing in parallel with automatic failover, while a frontier model handled only planning and synthesis.

A rate-limited provider hands its subtask on rather than killing the run, and swapping the crew is a config edit, not a code change.

Multi-AI Toolkit

Runs a homelab chatbot over offline Wikipedia and Arch Wiki snapshots using FastAPI, Ollama with Llama 3.2, and Streamlit — retrieval and generation both local, with the internet unplugged.

No API key, no per-token bill and no data leaving the box: the whole retrieval and inference path is self-hosted.

Zim Assistant

Works daily across Claude Code, the Claude API, Cursor, GitHub Copilot, ChatGPT / OpenAI API, Google Gemini, Ollama, Llama, DeepSeek, Groq, OpenRouter and Hugging Face — and builds the tooling those agents consume rather than only consuming it.

Prompt engineering and LLM API integration treated as an engineering surface with budgets and gates, not a chat window.

repo2graph