hey, i'm will.
I study math & CS at the University of Chicago. I'm an AI Security Research Fellow at XLab and I'm building Caisson AI, offline AI assistants for field technicians. Previously, I researched LLM watermark detection at MIT Lincoln Laboratory.
research
IEEE International Conference on Data Mining (ICDM) 2025 · Rising Star Award
Signature vs. Substance: Evaluating the Balance of Adversarial Resistance and Linguistic Quality in Watermarking Large Language Models
William Guo, Adaku Uchendu, Ana Smith
How well do LLM watermarks survive paraphrase and back-translation attacks, and what do they cost in linguistic quality? A systematic evaluation across watermarking schemes, from work at MIT Lincoln Laboratory.
experience
- U
UChicago XLab
AI Security Research Fellow
- Selected from 300+ applicants for a funded fellowship in AI agent security; built an LLM red-team agent that evaded provenance-based intrusion detection on 85 of 100 tasks.
- C
Caisson AI
Founder
- Offline-first RAG assistant over equipment manuals for field technicians: 95% retrieval accuracy, $50K across 3 paid pilots with elevator and heavy-equipment firms.
- Q
Quasi AI
Software Engineering Intern
- Built distributed simulation services with checkpointing and auto-retries (job success 88% to 99%); PyTorch DDP training cut runs from 40 to 17 hours at equal accuracy.
- C
Chicago Human+AI Lab
Undergraduate Researcher
- Prototyping multi-agent LLM pipelines that automate hypothesis generation, ranking, and evaluation over research literature.
- C
Calverton Growth
Private Equity Intern
- Deal sourcing and financial due diligence on manufacturing acquisitions.
- M
MIT Lincoln Laboratory
Machine Learning Research Intern
- First-authored an IEEE ICDM 2025 Rising Star Award paper on LLM watermark detection: benchmarked 4 schemes across 10K+ prompts and showed back-translation drops KGW detection AUC from 0.95 to 0.52; cut full-run compute from 48 to 6 hours.
- N
Northwestern University
Research Intern
- Built large-scale web scraping tools and applied ML clustering to analyze education data across Chicago Public Schools.
selected projects
view all
Caisson AI
Offline-first RAG over equipment manuals for field technicians: 95% retrieval accuracy with fully on-device search, in 3 paid pilots.
writing
view allMonitors that watch an AI agent's session miss harm that fires after it ends
Monitors that score an AI agent's commands as it runs them caught none of four attacks that hide the harm in a scheduled job or system trigger. Listing and resolving those triggers at session end caught all four, with no false alarms on four honest sessions.
ai safetyai controlmonitoringOctober 14, 20268 min readWhat Broke When I Ran Thirteen Claude Code Sessions on One Always-On Machine
Long-running agent sessions failed where they acted on a state they could not see: an account balance, an approval, another session's progress, or their own compacted context. Here is each failure and the change that fixed it.
agentsworkflowclaude codeOctober 14, 20266 min read
contact
The fastest way to reach me is email: wguo4@uchicago.edu. For anything longer, there's a form.