Selected Projects
Each card opens a write-up — the problem, the architecture, and the parts that were actually hard.
Unstuck
When a browser agent gets stuck on a challenge — a reCAPTCHA, a slider — it calls solve(page). A person claims the task and drives that browser from a phone until the page is clear. Built for a Solana hackathon, where a solved task cost 0.01 USDC from a prepaid balance. The public demo is the handoff alone, with no wallet. Packaged SDKs in TypeScript, Python, and Go are the next piece of work, and they have not been started.
kvroute
kvroute is a KV-cache-aware router in front of six single-GPU vLLM replicas, built to answer one question: where does cache-aware routing stop paying against plain consistent-hash session affinity? Six routing policies were built and five measured across 746 recorded cells — a working-set-by-skew pressure grid, a replica kill, and a per-request belief-divergence run. Cache-aware routing wins by +66.2% at the frozen comparison point and up to +373% under skew, but the two affinity policies' cache hit rates differ by at most 2.2 points, so the win is load balancing, not extra cache hits — a stateless hash of the prompt buys most of the same gain for a fraction of the machinery. Several of the project's own earlier headlines, including a six-card comparison and an exact-residency policy built to beat the router's belief, were measured and then withdrawn.
Yapper
An AI platform that turns any movie or video clip into a punchy funny-commentary track (短片解說 style) — condensed footage plays under an AI voiceover that retells and riffs on the story. A 13-stage resumable pipeline runs ASR, scene detection, multimodal LLM scripting, and voice-cloned TTS, orchestrated across dedicated GPU and CPU worker lanes.
LLM4Subjects
My first serious training project, rebuilt two years later. SemEval-2025 Task 5 asks for GND subject headings on bilingual library records, from a vocabulary of 79,427 candidates — and 19.1% of the headings the test split actually uses never appear in training at all. The 2024 version was a closed-vocabulary classifier, so it could not represent a fifth of the answer. The rebuild is retrieval: a neighbour harvest, a label-text tower and BM25, fused by reciprocal rank, then a LoRA-tuned encoder over a 70,579-document index.
HubStream
A globally distributed WebRTC Selective Forwarding Unit (SFU) platform for scalable video conferencing. Features real-time mesh routing and low-latency streaming.
Video Semantic Search
An end-to-end distributed system for querying video content using natural language. Features a Go backend, Next.js frontend, and Python ML infrastructure with Whisper and Qwen3-VL.
K8S Observability Platform
Built a comprehensive monitoring stack for a WebRTC application, implementing metrics collection, visualization, and alerting to ensure high availability and performance.