I like building AI agents — and then trying to break them. 🧪
My recent work focuses on runtime reliability, safe tool execution, recovery, memory, and evaluation.
- reliable Agent runtimes: state, recovery, validation, and memory
- safe tool use and human-in-the-loop execution
- evaluation that catches “looks successful, actually wrong” behavior
- retrieval / search systems when I need a break from Agent runtimes
A production-oriented Agent runtime for reliable multi-step execution.
It includes typed Planner decisions, policy-governed tools, durable execution, restart recovery, multi-tenant authorization, and governed cross-thread memory.
A result I particularly like: removing the Validator increased nominal completion from 50% to 75% — because invalid plans were being counted as successful. The runtime treats visible failure as better than silent success.
Built with: Python · FastAPI · Pydantic · SQLite · Docker
A safety-focused Agent/DataOps prototype that turns an Alertmanager incident into an evidence-gated, human-approved, restart-safe Kubernetes rollback.
The rule is simple:
The LLM may investigate and propose. It never gets authority to mutate infrastructure.
The workflow combines durable incident state, Prometheus evidence, resumable approval, allowlisted Kubernetes mutation, restart reconciliation, and post-rollback recovery verification.
The repository also includes a reproducible safety matrix on disposable kind clusters covering denial, replay, RBAC rejection, restart ambiguity, recovery failure, and cleanup.
Built with: TypeScript · Node.js · Prometheus · Kubernetes · OpenClaw · Vitest
Agent systems are where I spend most of my time lately, but I also enjoy:
- RAG and retrieval evaluation
- search and recommendation systems
- multimodal retrieval
- LLM application infrastructure
- ML systems where the interesting part starts after
model.predict()
Retrieval systems, recommendation systems, multimodal search, NLP experiments, quantitative research agents, and a probably unreasonable number of small ML projects.
A few repos worth browsing:
I like systems where failures are observable, decisions are inspectable, and “the demo worked once” is not the definition of reliability.
If a system can retry, recover, explain why it acted, and refuse unsafe work, I'm probably interested in it.