AI-agent systems + low-latency C++
CS @ City University of Hong Kong · Class of 2028 · Open to Summer 2027 SWE and AI-agent internships
I build AI-agent tools and performance-critical software, with an emphasis on measurable behavior, reproducibility, and failure handling.
orderbook — Low-latency C++23 matching engine
A price-time-priority limit order book and matching engine validated against real ITCH 5.0 market data, with an OUCH 4.2 order-entry gateway.
16.6–18.5 ns/op mixed · 11M+ differential checks · zero violations on full NASDAQ ITCH replay
C++23 · CMake · ASan/UBSan · ITCH/OUCH
Repository · Architecture and benchmarks
rof — Local Rust agent runtime + eval lab
A self-contained Rust agent harness: one loop, six policy-gated tools, hard budgets as the kill switch, and a write-ahead log where the log is the truth and the prompt is a view.
v1 (frozen as v1-frozen): 20-task suite averaged 17/20 over 3 reps (DeepSeek V4.1 Flash) · 5.1× fewer billed tokens/task than hermes at equal 18/18 · honest failure edge reported (large-deletion tasks)
v2 (merged to master): 14-crate workspace · frozen, digest-pinned Terminal-Bench slices graded from task-written reward files — the rig caught its own grader bug (a "25/26" that was really 6/26) · $0.006–$0.13 per eval cell vs $6–$23/task reported by frontier-tier harnesses
Rust · eval harness design · budgeted context assembly
roof-bot — WhatsApp AI sales agent
A Russian-language sales agent that qualifies leads, quotes from live pricing data, transcribes voice messages, and exports confirmed orders to Google Sheets and the owner's WhatsApp.
96% grounded · 92% correct on a 24-case evaluation set
Python · FastAPI · Gemini · LangGraph · RAG
PMOS — Multi-agent development framework
A host-portable framework for role-specialized agents with hybrid knowledge bases, human approval gates, spend tracking, and artifact traceability. Runs on Claude Code, jcode, Hermes, and OpenHands.
Used by a PMOS agent team to ship cronx, a DST-aware, standard-library-only cron explainer.
poKING — seeded PPO and NFSP research in a poker engine, built around honest measurement. Includes a write-up of a benchmark bug that inflated measured variance by approximately 2,800×.
- Core: Python · C++ · Rust · TypeScript/JavaScript
- AI and agents: Gemini · LangGraph · RAG · evaluation harnesses
- Systems and product: CMake · FastAPI · SQLite · Next.js · Firebase
English C1 (IELTS 7.5) · Russian · Kazakh