Russell Ro

ML researcher. PhD candidate @ UC Berkeley Bioengineering, with a major emphasis in Computer Science. Advised by Reza Abbasi-Asl. NSF GRFP fellow.

I research and build ML and agentic systems. Currently focused on:

• agentic workflows and search/retrieval for clinical trial design,
• representation learning on model zoos for sample-efficient LLM profiling and model selection (distillation detection next),
• and production agentic systems for internal operations, plus causal inference harnesses for client-facing services, as an ML Engineer (consultant).

News:

10/2026: Our paper “From the Zoo to the BULLPEN: Sample-Efficient LLM Profiling via Multi-Channel Model Embeddings” is under review at ICLR 2027.

10/01/2026: I gave a talk, “eval ∝ d: evaluating evals before they dictate deployment decisions”, at an AI Decisioning Working Group, a small cross-disciplinary meetup of tech professionals. Slides (PDF)

07/26/2026: Our paper RECITE: Revising Eligibility Criteria Incorporating Textual Evidence was accepted to KDD 2026 (AI4Sciences track). I presented it in Jeju in August. RECITE is a benchmark for evaluating LLMs on revising clinical trial eligibility criteria using evidence from protocol amendments.

Currently building:

• Evidence-guided clinical trial design: LLM pipelines and retrieval over 500K+ trials for eligibility criteria analysis and literature–trial matching, plus large-scale studies of how criteria evolve and whether trials adopt published recommendations
• BULLPEN: multi-channel model-zoo embeddings for sample-efficient LLM profiling (under review at ICLR 2027)
• Multi-agent orchestration for industry and research: a framework of messaging-interfaced agents with multi-model routing and subagent deployment that runs my research operations day-to-day

Other things I’ve done:

• Adversarial robustness testing of medical LLMs (arXiv)
• Designed a high-throughput drug screening system at Novartis (article)
• ScanAlert: electronic medication adherence monitor
• Satellite test engineering at Amazon (Project Kuiper)

Honors: NSF GRFP, Jacobs Scholar, Tau Beta Pi, Intel ISEF Finalist

For fun:

• Custom game controllers: toy wheels, toasters, and webcam pose estimation turned into Xbox gamepads and keyboard inputs (Arduino, Python, MediaPipe). Up to 4-player local multiplayer. Demoed at Open Sauce 2026.