Distributed backend systems · LLM agent infrastructure
M.S. Computer Engineering @ Columbia
Dec 2026
B.E. ECE @ UMich–SJTU JI
I work on the unglamorous half of systems: the execution layer underneath, the part that has to stay correct when things run concurrently and for a long time.
Right now I'm a SWE intern on TikTok Trust & Safety — Algorithm Infra, on the async execution layer behind a safety-model evaluation platform. I re-architected a synchronous SSE path into MongoDB-backed durable tasks — a 15-minute ceiling became 3-hour jobs — and extended a Go DAG engine with async triggers, which is how I ended up tracking a production concurrency deadlock across parallel SG/US runs.
|
Interested in
|
Looking for
|
| Project | The hard part | Stack | |
|---|---|---|---|
| ⚡ | mbook | Deterministic limit order book. O(1) dense tick ladder over a preallocated slab, nothing allocated on the hot path, replay verified by a digest over the ordered fill stream. | C++17 CMake |
| 🔒 | An-agent (private) | Multi-agent assistant on the Claude Code SDK. Five-layer memory (procedural → working → episodic → semantic → persona), and a read-only digital twin with three-layer prompt-injection defense. | Python SQLite/FTS5 ChromaDB |
| 🧭 | cs-interview-coach | Local-first, sub-millisecond FST search over 140+ knowledge cards, with FSRS spaced repetition across five navigation views. | Rust Tauri SvelteKit |
| 📮 | gmail-digest | Rules-first triage across several inboxes. The optional LLM pass must quote the email for every judgment it makes — no unsourced verdicts. | Python Discord API |
| 🎬 | xhs-bilibili-distill | One distillation pipeline, two source adapters: transcript + scene-detected key frames + structured report. | Python FFmpeg |
| 📊 | portfolio-insights-generator | Transaction analytics with LLM-assisted insight generation and what-if scenario modeling. | Python Streamlit pandas |