YuhaoCheng.
ML systems & research engineer — M.S. Computer Science, UIUC
Scroll to beginI build the systems underneath models — the pipelines that train them, serve them, and measure what they can actually do.
Seven sets below, one for each chapter: a campus, an evaluation pipeline, a GPU cluster, a dashboard studio, an inference farm, a benchmark gallery and a retrieval funnel.
- 0M+
- LLM calls orchestrated in one evaluation run
- 0d → 21h
- Full benchmark run, after load-testing the pipeline
- 0 img/min
- SDXL throughput across 8 GPUs, up from 6
- 0+
- WebArena tasks for SFT → RL web agents
Education
University of Illinois Urbana-Champaign
Aug 2025 — Dec 2026
Master of Computer Science
Aug 2021 — May 2025
B.S. in Computer Engineering
Research
TIMAN Group, UIUC
Research Assistant · May 2026 — Sep 2026
Auditing large language models at scale.
- Built a Python evaluation pipeline that executed 3M+ LLM calls across 8 models on a 64-worker pool, with a token-bucket rate limiter and exponential-backoff retries.
- Load-tested the model endpoints to their limit, raising throughput from 40 to 2,400 calls/min — a full run went from an estimated 52 days to 21 hours.
- Designed BDSU audits of GPT-4o, GPT-4o-mini and o1, decomposing behavior into demographic disparity, prompt sensitivity and generation uncertainty.
- Made runs resumable with request-level caching and checkpointing; every generation, score and metric is tracked in Weights & Biases.
Research
IBM–Illinois Discovery Accelerator Institute
Research Intern · Jul 2024 — Dec 2024
Teaching an 8B model to browse the web.
- Built an SFT-to-RL pipeline for Llama-3.1-8B web agents over 800+ WebArena tasks: browser interaction, trajectory collection and automated evaluation.
- Implemented actor-critic RL with dense trajectory rewards — task success, progress, URL similarity, exploration bonuses, state-change penalties.
- Scaled full-parameter training across 4 nodes / 16 GPUs with PyTorch, Accelerate and DeepSpeed ZeRO-3; 16 parallel browser workers cut an eval pass from 20 h to 3 h.
Industry
visibilityx.ai
Frontend Developer Intern · Jun 2025 — Aug 2025
Data-heavy dashboards that load fast.
- Built a Vue 3 + TypeScript SPA with 10+ data-intensive dashboard views; extracted 12 shared components and composables, removing ~2,000 lines of duplicated logic.
- Cut median dashboard load from 2.4 s to 1.6 s with parallel REST calls, Pinia response caching and per-route lazy-loading of ECharts.
- Wrote 180+ Vitest unit tests (80% line coverage), run in CI on every pull request.
Industry
HiABR Lab
Backend Developer Intern · May 2024 — Aug 2024
Serving Stable Diffusion XL on a GPU cluster.
- Deployed SDXL as a multi-node inference service — 2 nodes, 8 GPUs, one FastAPI worker per GPU — behind Nginx least-connections routing with health checks.
- Traced OOM crashes to overlapping requests on 24 GB GPUs; fixed them with bounded per-GPU job queues, then added dynamic micro-batching (4 prompts / 50 ms window).
- Under Locust load, went from failing above 4 concurrent users to 64 with zero errors; throughput scaled from 6 to 140 images/min at 21 s p95.
- Built a URL-shortening service on FastAPI, PostgreSQL and Redis with sliding-window rate limiting, idempotency keys and Prometheus metrics.
EMNLP 2026
VGI-Bench
Probing Visual Intelligence in Video Generation Models
- Co-developed a visual-reasoning benchmark across four task domains and seven capability dimensions.
- Built pipelines automating inference, collection and analysis for 9 video and 11 image generation models; studied failure modes, input sensitivity and reasoning dynamics.
2026 — ongoing
PIR-Arena
Proactive Information Recommendation Benchmark
- A multimodal benchmark for deciding when and what information to surface from continuous user context.
- Retrieval pipeline linking LLM query generation with BM25, dense retrieval and cross-encoder reranking over 50k documents at 1.8 s p95; evaluated on precision, recall, relevance and timeliness.
01 Languages
- Python
- TypeScript
- JavaScript
- C / C++
- SQL
- Java
- Bash
02 ML Systems
- PyTorch
- DeepSpeed
- Accelerate
- SLURM
- Weights & Biases
03 Backend & Infra
- FastAPI
- PostgreSQL
- Redis
- Docker
- Nginx
- Prometheus
- Locust
- AWS
- GCP
04 Frontend
- React
- Vue 3
- Node.js
- Express
- ECharts
- Vitest
Let’s make the next scene together.
Open to ML systems, infrastructure and research engineering roles — from Dec 2026.
yuhaoc7@outlook.com