AI engineer. I build agentic systems and the infrastructure to trust them: permission-aware retrieval, agent observability, human-in-the-loop approval, and evals that fail CI when behaviour regresses. Building continuously.
Open source: Google Summer of Code, Linux Foundation (LFX), and Symmetry Autumn of Code. 3+ years shipping in production.
Agent infrastructure
- vaultrag: Permission-aware RAG: access control enforced inside the retrieval query, CI fails on any leak.
- agentrace: Observability for Claude Code subagents: flags the agent results you should not trust.
- voiceeval: Evaluation for voice agents: catches the failures a text eval scores as a perfect call.
- Tracecase: CI for AI agents: replay runs against prompt changes, catch regressions before they ship.
- Bridgekit: Scoped MCP server with per-client permission boundaries and an append-only audit log.
- casebook-mcp: Remote MCP server over AgentPostmortem, a public registry of real AI-agent failures.
- casebook-chat: Streaming chat that investigates agent failures and answers with cited case IDs.
Agents in production
- Resolvd: End-to-end inbox operator: auto-resolves within policy, escalates the rest with a proposed action.
- Greenlite: Approve your agents from your phone: one-tap human-in-the-loop for escalations.
- Webhands: Computer-use agent for tools with no API: refuses any write unless you confirm it.
- relayg: Support triage as a LangGraph state machine, paused mid-run by a human-in-the-loop interrupt.
- AgentPostmortem: Every AI agent failure, documented: a public case registry.
Products
- SignalizeAI: AI prospecting as a Chrome extension, featured on the Chrome Web Store.
- PennyRush: A private money hub: receipts and statements parsed in memory, only extracted fields stored.
- ProofClip: Testimonial walls and social-proof cards for creators and small SaaS, embedded with one script tag.
Roots
- cpp: The code I wrote while learning C++ at the very start, kept detailed on purpose: datatypes through graphs and DP.