Agent workflows
Building typed tool workflows that trace proposed state changes to bounded evidence and keep risky actions behind human review.
Agent Systems, LLM Infrastructure, and AI Reliability
Engineering Focus
Retrieval, memory, and evaluation
Building provenance, freshness checks, failure analysis, deterministic eval contracts, and privacy boundaries around model and retrieval behavior.
Model serving infrastructure
Building request harnesses, comparable artifacts, TTFT and throughput reports, KV-cache estimates, telemetry, and claim gates for serving experiments.
Relevant AI Projects
Computer Use Runtime
A computer-use runtime that compiles one model-driven exploration of a no-API interface into a typed, versioned capability, then replays it deterministically with no model in the loop.
careeroc
A career workspace that combines versioned resumes, recruiting history, and user corrections into evidence-backed recommendations, with human review before any update lands.
Boring Agent Memory
A canonical-first local recall layer using SQLite FTS5/BM25 over trusted files, with provenance, snippets, freshness checks, redaction, and workspace filtering.
LLM Serving Benchmark Harness
An OpenAI-compatible serving benchmark that records TTFT, inter-token latency, throughput, and SLO goodput, then audits its own output by recomputing every published metric from raw request rows.
Agent Sandbox Eval
A Docker-sandboxed framework for reproducible tool-using agent evaluation with task manifests, JSONL trajectories, deterministic grading, replay, and failure analysis.
Explore More
For broader engineering work, review the full project list or the cloud and data systems page. For opportunities, contact Hanbin directly.