Lead Software Engineer at a NASDAQ-listed SaaS group: I own service agreements end to end (it drives the product's ~$8B a year in customer payments), led the engineering of a 600-table, 21 TB MySQL core's decomposition into 20+ services, and built the AI platform ~20 engineers ship on daily, under one rule: never merge code you can't explain.
What's here:
- agentweft - agent harness for software delivery: controls as code (gates, per-step tool grants checked on every output, call and token budgets), journalled and resumable runs, MCP client and server; open-sourced from the workflow ~20 engineers use daily, hundreds of billions of tokens through it since May 2025.
- warpmap - Go audit CLI that maps a codebase for humans and coding agents: dependency graph, churn x complexity hotspots, blast radius, a baseline diff, a GitHub action that fails a risky PR, an MCP server.
- tare - eval harness for LLM output: its deterministic rules benchmarked for precision and recall on a hand-labelled set behind a regression gate; its exit code gates CI.
- react-formkit - define a form once as config and render it through an adapter for any React UI library, on React Hook Form and Zod; extracted from production work.
- weftworks, efquerykit - Claude Code skills and workflow patterns, an EF Core query kit for MySQL.
Upstream: 100+ merged pull requests, in Anthropic and OpenAI repos, Microsoft's onnxscript and Microsoft.Extensions.AI, NVIDIA's agent toolkit, Hugging Face, docling, deepset's haystack, vite, react-router and a dozen more.
Writing: HackerNoon articles and recorded webinars, from system architecture to agentic development:
- From idea validation to scale: how to choose the right architecture
- Turning a messy codebase into something you can actually work with
- Legacy in months, not years: keeping AI-written code safe to change
Hackathon judge: AI Slop Scan 2026 and DOGFOOD 2026 by Hackathon Raptors.
Happy to connect on LinkedIn.
