Hi, I'm Yincheng Zhou
CS & Physics @ UPenn VIPER
Building intelligent systems that work reliably at scale.
I am a dual-degree student at the University of Pennsylvania in the VIPER program, studying Computer Science (AI concentration) and Physics & Astronomy. I build intelligent systems across the full stack — from training vision-language models to engineering production infrastructure that scales.
Right now, I am working on two research fronts: at UPenn's GRASP Laboratory, I proposed SPD, a sparse monocular depth estimator achieving 6x inference speedup for real-time robotics; and at Peking University's HMI Lab, I built distributed training infrastructure for GUI agents across 112 Docker VMs and 8 A100 GPUs.
Previously, as the Founding Software Engineer at Franklink, I architected a multi-agent platform from zero to one with Kafka-backed event pipelines sustaining 50+ concurrent sessions on AWS ECS.
Proposed SPD, a sparse monocular depth estimator that predicts depth only at queried pixels, achieving 6x inference speedup over dense methods while preserving accuracy (AbsRel 0.11 on NYU Depth V2). Developed a dense-train / sparse-infer paradigm with a 350K-param split decoder and cross-attention fusion on DINOv2 features. Advised by Prof. Pratik Chaudhari.
Developed MCTS-guided SFT for GUI agent training with step-level decisions and difficulty weighting, outperforming trajectory-level RL (GRPO) at 25% compute and raising UI-TARS-7B's OSWorld success from 12.3% to 17.7%. Engineered distributed rollout infrastructure across 112 Docker VMs and 8 A100 GPUs.
Designed iMessage-based multi-agent orchestration with persisted memory, selective dispatch across 5 sub-agents, and agentic RAG over 8 permissioned tools. Engineered a Kafka-backed event pipeline with idempotency keys and retry + DLQ, sustaining 50+ concurrent sessions on AWS ECS.
iMessage-based multi-agent orchestration with selective dispatch across 5 sub-agents and agentic RAG over 8 permissioned tools, backed by a Kafka event pipeline sustaining 50+ concurrent sessions on AWS ECS.
Per-ticker financial sentiment platform powered by a DeBERTa model fine-tuned to 82.5% accuracy, with a 3-pass NLP pipeline ingesting from 7 subreddits through concurrent-safe batch processing.
Hackathon-winning civic platform that converts citizen grievances into ready-to-send advocacy with root-cause analysis and elected-official mapping across 4 government levels.
PPO agent trained across 80 parallel SNES emulators with spline-interpolated reward shaping, reaching competitive performance on a task previously unsolvable due to sparse 3,000+ step horizons.
Multi-agent incident remediation system that simulates outcomes before execution to reject unsafe plans, achieving 46% faster recovery with a closed-loop plan-execute-verify pipeline.
5-agent orchestration on Google ADK converting natural language to validated SQL with loop-guarded retry, 18 security patterns, and automated 6-type chart generation across 3 SQL dialects.