Research & engineering
Systems I have built to study agents, evaluate their behavior, and support real workflows.
FarField
Python 3.11 · standard library
A reproducible research loop, from a question to an experiment.
- Problem
- Turning a broad topic into a falsifiable hypothesis and a registered, two-arm experiment protocol.
- Implementation
- Built a local workflow in which the model proposes claims and scripts, while graph checks, prior-art checks, dataset binding, and experimental arithmetic remain ordinary Python.
OpenClaw Security Suite
TypeScript · Node.js · Vitest
Security infrastructure for a tool-using agent gateway.
- Problem
- Making agent actions inspectable and skill execution subject to explicit policies.
- Implementation
- A persistent audit-event store with signed integrity checks; rule-based alerts across webhook, Slack, Discord, and Telegram; a configurable skills sandbox; and a security dashboard with eleven permission-scoped RPC methods.
IntelliSelect
FastAPI · React 19 · vLLM · Docker
An auditable assistant for faculty recruitment pre-screening at HKUST (Guangzhou).
- Problem
- Combining evidence from multiple sources while keeping recruitment decisions with people.
- Implementation
- Integrated five-dimension scoring with a self-hosted Qwen3-14B model, deterministic matching rules, and Google Scholar peer-group benchmarking in a full-stack application.
Legal Pilot 2.0
Qwen3-235B · FastAPI
A two-stage workflow for reviewing contracts.
- Problem
- Locating legal risks and suggesting revisions while protecting sensitive information.
- Implementation
- Combined a legal red-line scan with rule-by-rule auditing, then added sensitive-entity detection, deterministic pre-masking, and length-preserving replacement.
Agentic RAG & Memory Systems
PyTorch · Hugging Face · Faiss · Milvus
Evaluation and data pipelines for agents working across long contexts.
- Problem
- Understanding information-recall failures, context drift, and task-level memory needs.
- Implementation
- Built a long-context, cross-topic retrieval benchmark and enhancement pipeline, and converted browser-agent trajectories into task-level SFT data.