Projects & writing
Selected work
Open-source tools and applied engineering — plus selected published writing.
Open source
Assevra AI
An MIT-licensed, open-source reliability scorecard for LLM agents. Assevra AI scores agent reliability against fixed thresholds so teams can measure — not guess — whether an agent is safe to ship.
Founder and maintainer.
- Scores agent reliability across four dimensions: grounding and faithfulness, safety and refusal, PII-leak detection, and task completion.
- Evaluates against fixed thresholds with 95% Wilson confidence intervals for statistically honest pass/fail signals.
- Ships as a CLI with Markdown, JSON, and HTML reports plus a CI gate for continuous reliability checks.
- Written in Python and published as an installable package on PyPI.
Published writing
- HackerNoon
Claude Code Doesn't Ship Production Code: Your Config Does
Why the configuration around an AI coding agent — not the model — determines whether it ships production-ready code.
- HackerNoon
Stop Letting Your AI Agents Call Each Other Directly
The case for mediated, governed communication between agents instead of brittle direct agent-to-agent calls.
- Platform Engineering
Your Newest Platform User Is an AI Agent — Build It a Golden Path
Treating AI agents as first-class platform users, with golden paths, guardrails, and paved-road tooling.
- DevOps.com
Multi-Model AI Debate: The New Way to Design Software Systems
Using structured multi-model debate to pressure-test architecture decisions and design software systems.
- Security Boulevard
Your AI Agent Is an Insider Threat With a Shell
Why autonomous agents with tool access should be governed like privileged insiders — and how to contain them.
- BizTech Foundation
AI Agents Don't Have an Intelligence Problem. They Have a Bookkeeping One.
The reliability bottleneck for enterprise agents is state, memory, and auditability — not raw model intelligence.