Research
MTEB-Gym: label-free embedding evaluation
Research collaborator | Jun 2026 - Present
How do you choose an embedding model when your dataset has no human relevance labels? I co-developed MTEB-Gym to investigate that question, then designed a judge-reliability study across 16 retrieval datasets to find out when an LLM-based ranking could actually be trusted.
The main failure mode was task-definition mismatch. Giving an evaluator the wrong idea of what a dataset was asking could change its judgment. I worked on dataset-specific criteria and built a public leaderboard that reports reliability alongside the model rankings.
Nine merged upstream changes cover validation, uncertainty estimates, cache behavior, deterministic parallelism, and failure handling. The reliability study is still in progress. This is part of my work in the wider MTEB ecosystem, whose package had reached roughly 17 million PyPI downloads by August 2026.
RSNA Knee Abnormality Detection
Independent researcher | Jul 2026 - Aug 2026
I built a knee-MRI research pipeline for the Radiological Society of North America's Knee Abnormality Detection competition. It predicts 12 study-level findings across 4,407 studies and 24,371 imaging series.
The system includes 2.5D ResNet-18 MIL, five-fold and three-plane EfficientNet-B3 training, DINO and RadImageNet branches, rank fusion, and label-specific fallback policies. I built reproducible offline execution, geometrically ordered slices, hash-bound caches, checkpoint recovery, and inference replay. These are competition and public-leaderboard results, not a clinical deployment.
AgentBreed
Sole author | Mar 2026 - Jul 2026
I wanted to know how much multi-agent performance comes from the search algorithm and how much comes from the set of choices it is allowed to search. AgentBreed was my preregistered attempt to separate the two.
I ran 700 reproducible evaluations across three domains in a deterministic simulator, using equivalence testing and sensitivity analysis. The tested search operators did not show Holm-corrected significant pairwise differences in that pilot. The result belongs to that simulator; I built the study to make the distinction inspectable.
Procedural-Maze RL Baselines
Sole author | Feb 2026 - Apr 2026
Why could a five-line heuristic see a policy that PPO, DQN, and A2C kept missing? I built a procedural-maze benchmark to compare those agents with simple heuristics and behavior cloning.
The models had enough capacity. Exploration and credit assignment were the problem. I kept roughly 3,500 traceable run artifacts, found a flaw in my original validation setup, and rebuilt the evaluation around a proper held-out protocol. Correcting the experiment became part of the project.
Connectome Architecture Benchmark
Independent researcher and auditor | Dec 2025 - Aug 2026
Biological wiring is an appealing starting point for a neural network. I wanted to see what remained of that appeal after controlling for density, weights, graph realizations, and trainable components.
I benchmarked ten connectomes across six tasks, kept the biological wiring fixed, and trained only the output layer. Biological structure sometimes helped, but degree distribution explained much of the advantage; the exact wiring explained less. The useful part was finding out which part of the story survived the controls.
Regional Cabled Array expedition RR2607
Student researcher and pipeline developer | Aug 2026 - Present
In August 2026, I joined the Regional Cabled Array expedition RR2607 through VISIONS '26. I worked with WHOI, NSF's Ocean Observatories Initiative, and UW APL, deployed subsea instrumentation, and surveyed Axial Seamount, about 300 miles off Oregon, using the 11,000-pound Jason ROV.
My ongoing research asks whether methane-bubble flux can be estimated while the experiment is 2,900 meters underwater. I'm building a pipeline that brings together ROV video, sonar, hydrophone, CTD, and navigation data. The shipboard sources stay read-only, and every derived artifact traces back to source paths, hashes, and code versions.
The computer vision, acoustic alignment, calibration, and held-out validation are still in progress. The expedition happened; the final methane-flux result is still being worked on.
Technical notes
Technical report, advised by Professor Zhuang Liu at Princeton University.
I explored an AlphaEvolve-style search over prompts, parsing, formatting, and inference code for visual chart questions. MAP-Elites kept the candidates in a behavioral archive, with an inspectable history for each accepted mutation.
The best single-prompt Qwen3-VL configuration reached 63.25% held-out accuracy on CharXiv, but its difference from the 60.5% manual prompt was not significant (p = .42). A separately designed per-question router reached 73.0% against a 29.5% naive baseline, with a paired p-value of 2.4e-6. Those are separate comparisons, and the report keeps them separate.
Technical report, advised by Professor Juncheng Yang at Harvard University.
I built an 80-point vLLM serving matrix to compare dense and mixture-of-experts models on one H100, at roughly equal active-parameter budgets. The sweep covered prompt length, generation length, concurrency, and serving configuration for Qwen3.5-9B and GLM-4.7-Flash.
The throughput crossover appeared near concurrency three. At 49K-token generation, Qwen admitted 42 concurrent requests to GLM's 10, and GPQA accuracy moved by 20 points with token budget. A hardware audit also caught a PCIe-throttled H100 before publication. The machine running the benchmark deserved as much attention as the models.
Published work
Three peer-reviewed papers, 57 citations, and an h-index of 3 as of 2026-09.
- Design of systems for plasma activated water (PAW) for agri-food applications, Journal of Physics D: Applied Physics, 2024
- Design of a continuous plasma activated water (PAW) disinfection system for fresh produce industry, Innovative Food Science & Emerging Technologies, 2024
- Design and construction of a continuous industrial scale cold plasma equipment for fresh produce industry, Innovative Food Science & Emerging Technologies, 2024
Other things I built
Mimic lets me teleoperate a simulated Franka arm from a browser and work on imitation learning. It brings robotics, computer vision, and learning from demonstrations into a system I can interact with.
Engram explores online learning through local plasticity and persistent memory. I built it as a prototype for experimenting with how a system might keep learning over time. Its experimental efficacy is still unverified.
Forge is my provider-agnostic agent runtime. It brings tools, memory, budgets, and observability into one place so I can work on how agents run and inspect what they do.
WireML is a terminal workbench for training lightweight heads on foundation-model embeddings. It is a way to work directly with representations and small models without leaving the terminal.
delphi-quant is a retrospective verification harness for systematic strategies. I built it around reproducibility and making rejection legible: a failed strategy should leave behind an understandable reason it failed.