Boqin Yuan
Building the evaluation, post-training, and serving infrastructure that makes long-horizon LLM agents work
b4yuan[at]ucsd.edu
I’m a Master’s student in Computer Science at UC San Diego, where I work with Prof. Jishen Zhao’s STABLE Lab on agent memory and ML systems. I spent summer 2026 as a Software Engineer Intern (AI Agent) at Moody’s Analytics in San Francisco, working on knowledge iteration and evaluation for their banking agent. Before UCSD, I was the founding Machine Learning Engineer at CambioML (YC S23) for over a year, where I trained and deployed AnyParser, a vision-language model for document parsing, and built Energent.ai, a computer-use agent sandbox. I studied Mathematics & Computer Science and Statistics at UIUC before that.
These days I spend my time on three things:
- Agent evaluation and benchmarks - benchmarks and diagnostics for long-horizon LLM agents: SkillsBench, Agents’ Last Exam, AMA-Bench (ICML 2026), and memory-probe. I also contribute to open-source agentic benchmarks: Harbor, Terminal-Bench, QF-Bench, and AgenticVBench.
- Model training and alignment - post-training LLMs and VLMs: supervised fine-tuning, LoRA, and RL alignment with PPO, DPO, and GRPO in PyTorch and DeepSpeed.
- ML systems and inference - serving models under production load with SGLang, vLLM, speculative decoding, quantization, Kubernetes, and AWS. Studied speculative tool calling for agent latency.
Graduating Dec 2026, open to full-time roles from Jan 2027.
Recent Experience
Industry
Software Engineer Intern - AI Agent
Moody's Analytics · Internship
Work on knowledge iteration and agent evaluation for Moody's banking decision-intelligence agent: injecting domain knowledge through skills, then measuring and optimizing against Langfuse traces, inside a LangGraph orchestration on AWS.
Founding Machine Learning Engineer
CambioML (YC S23) · Full-time
Trained and productionized AnyParser, a 1B & 2B vision-language model that parses PDFs, including tables and charts, into structured Markdown: sourced PDFs from Common Crawl, built the cleaning, filtering, and annotation pipeline, then fully fine-tuned and preference-aligned on 8×A100s, beating GPT-4 baselines on DocVQA (ANLS). Served with SGLang at 150 output tokens/s on a single L4 GPU, deployed as a SaaS on AWS (ECS + Lambda).
Core contributor to Energent.ai, a Claude-powered computer-use agent (CUA) sandbox with multi-agent orchestration, long-term memory, and per-user Kubernetes-isolated VM sessions, reaching 1,000+ registered users. Built a computer-use gym: a reproducible desktop environment with defined tasks where agent rollouts generate the trajectories used to evaluate and post-train CUA agents.
Whitepaper
- Achieving 2x Accuracy in Knowledge Retrieval from Charts and Tables without Intensive Prompt Engineering · CambioML & Epsilla, 2024
Machine Learning Engineer Intern
Inspur Group · Internship
Built and annotated a custom volleyball dataset, then trained YOLOv7 detectors and YOLO Pose keypoint models for real-time athletic movement analysis. Accelerated inference with TensorRT to cut latency while preserving accuracy.
Research
Research Assistant - STABLE Lab
UC San Diego · Prof. Jishen Zhao
Agents fail on long-horizon work for three separable reasons: the context they are served, what they retain from it, and how they were trained to act on it. ML systems: CodeNib compiles lexical, dense, and structural views of a repository once per commit and maintains them incrementally, so coding agents receive bounded context instead of rediscovering the codebase every task, with graph and vector updates running 8.7× and 25.4× faster than an independent rebuild at the median across 100 repository snapshots, on the subset where outputs match. Agent memory: AMA-Bench measures whether agents carry state across extended task sequences, rather than scoring the single-turn recall that most memory benchmarks reward. RL training: PRO-V-R1 post-trains an 8B programming agent for RTL verification with SFT plus GRPO on verification rewards derived from program-tool feedback, beating GPT-4o on VerilogEval-v2 functional correctness (57.7% vs 43.0%).
Publications
Research Assistant - SSAIL
Supercomputing System and AI Lab, UIUC · Prof. Minjia Zhang
Extended DeepSpeed-Chat, whose RLHF pipeline shipped with PPO only, to preference-based post-training. Built a DPO trainer that scores chosen and rejected responses against a reference policy with adaptive KL control, trainable with LoRA, added a best-of-N rejection-sampling variant, and extended the preference-data pipeline to feed both. Runs on DeepSpeed's hybrid engine and ZeRO.
Code
- boqiny/deepspeed-chat · DPO and rejection sampling for DeepSpeed-Chat
NCSA SPIN Research Intern
National Center for Supercomputing Applications · Prof. Sheng Wang
Machine learning for geospatial and remote-sensing data: multi-temporal crop classification with pre-trained vision foundation models, and field-level crop-residue and tillage-practice detection across the EU from multi-source satellite imagery.
Publications