M.S. Computer Engineering @ Columbia, graduating Dec 2026. SWE intern at TikTok Trust & Safety this past summer, now Lead Data Engineering & ML Intern at DiyoWater; previously Johnson & Johnson and iQIYI. I make LLM agents, ML systems, and vision pipelines measurable, debuggable, and reliable.
I'm a Computer Engineering master's student at Columbia University with six internships and research roles across TikTok, DiyoWater, Johnson & Johnson Medical, iQIYI, UAES, and SJTU. My sweet spot is the unglamorous last mile of AI: evaluation, reliability, and infrastructure, making agents measurable, debuggable, and safe to ship.
I've built LoRA fine-tuning fleets and TensorRT edge deployments, multi-agent daemons with five-layer memory systems, TLA+-verified distributed locks, and Go async task infrastructure for an LLM agent in production. I'm especially drawn to computer vision: detection on the edge, diffusion and LoRA, computational imaging, and 3D reconstruction. If it involves a model meeting the real world, I've probably debugged it at 2 a.m.
Off the clock I'm usually behind a camera, on a bouldering wall, mapping my travels into a hand-built interactive atlas, or drawing pixel worlds. I like building things on both sides of the art and engineering line.
Research on how to evaluate production conversational LLM agents: what goes wrong in real dialogues, which failures matter most, and how to close the loop without new live traffic.
Working across turn-level analysis, ranking of failure patterns, and user simulation on real customer-service conversations.
Details stay under wraps until publication.
Long-lived multi-agent daemon: PID locking, signal handling, launchd batch jobs, structured incident logs, shell-level command interception.
5-layer memory (SQLite FTS5 + ChromaDB): sub-100ms recall at 100K+ entries with daily GC and graceful degradation.
3-layer rate limiting + circuit breakers keep misbehaving agents inside API quota.
Three-stage YOLOv8 pipeline (hold detection, 5-class type, volumes): mAP50 0.825 on hold detection.
Exported to ONNX, quantized to TensorRT FP16 on Jetson Nano: 12 FPS single-stage, 5 FPS full pipeline at 5W.
A* beta solver with scale-invariant reach recommends optimal move sequences via a Flask MJPEG WebUI.
LLM agent orchestrates macro analysis, multi-factor scoring, RL portfolio optimization, and news, behind a schema + risk + anomaly validation layer.
Multi-model routing picks vendor and model size per task: under $15/month for weekly rebalancing over a 100-stock universe.
Every decision recorded as a structured audit log for offline post-mortems and prompt/policy refinement.
Lease-based distributed lock with fault-tolerant renewal across crash and message-loss failure models.
TLA+ spec model-checked with fault injection: mutual exclusion, lease validity, and monotonicity invariants proven.
Property-tested the Python implementation with Crosshair, surfacing race conditions in lease handoff.
Real-time multimodal pipeline: Pinyin parsing, error correction, ranking, and speech/image feedback under <1s end-to-end latency.
Trie retrieval + probabilistic correction + LLM fallback rerank reach >95% accuracy on noisy input.
Async LLM calls with non-blocking UI and fallbacks on resource-constrained hardware.
Anime characters living on the desktop: image to 3D mesh (Hunyuan3D), auto-rigged, walk cycles and spell animations rendered with three.js + three-vrm inside a transparent, always-on-top WKWebView.
They actually do things: calendar reminders via EventKit, a drag-and-drop "magic shelf" for files, clipboard capture, screen snipping, and face-based photo sorting from the right-click menu.
Five characters shipped as standalone .app bundles with a one-script build; 2D sprite-sheet and 3D variants share the same Swift host.
My U.S. trips as a magical interactive map: per-trip diaries, hand-drawn route animation, and venue markers reconstructed by running OCR over my own photos.
Build pipeline turns raw GPS, photos, and notes into a single static page; a public build strips every personal photo so the shareable version carries zero private data.
Zero frameworks at runtime, works offline, renders on a phone.
C++17: O(1) dense tick ladder over a preallocated slab, no hot-path allocation, replay-verified determinism.
Tauri + SvelteKit + Rust: sub-ms FST search, FSRS spaced repetition, 140+ cards across DSA/ML/LLM/RecSys.
Measures ten facial profile angles from a side photo, entirely client-side; shows population norms, uploads nothing.
Click-to-fall stress-relief game with combo scoring and depth-themed stages; runs as a live wallpaper via Lively.
Faces to studio-style group portraits via Gaussian splatting with relighting; environment and object stage validated on a 4090.
Natural language to robot insertion depth: 4-tier router, 100% on a 372-case evalset (ALOHA bimanual).
Several inboxes into one Discord feed; the optional LLM pass must quote the email for every judgment it makes.
One pipeline, two source adapters: transcript + scene-detected key frames + structured report.
Local-first fitness app: Claude/GPT-4V coach with 16+ tools, voice + image input, 13 toggleable modules.
Proof-of-work chain in Python: Ed25519 transactions, fork resolution, mempool recovery on reorg.
Spring Boot microservices on AWS ECS with CI/CD, MySQL/MongoDB, and CloudWatch observability.
Solidity + Hardhat contracts (no high-severity Slither findings, 100% test coverage) with Next.js + wagmi frontend.
I shoot on a Canon R6 Mark II, mostly travel and portraits. Here's a first set from Peru; the archive is big and I'm bringing it over slowly.
Open to 2027 new-grad and internship roles in software engineering, machine learning, computer vision, and agent systems. I reply fast.