01 · shared compute
Lab-scale AI infrastructure
GPU clusters, Slurm, IaC, monitoring, quotas, and the everyday question of how to make shared compute usable, fair, and reliable.
Research is a system.
I like building the system.
Compute, experiments, agents, representations, and lab workflows all shape what research is possible. I work on the pieces between them: cluster operations, reproducible research tooling, autonomous experimentation, and structure-aware AI models.
Interests
if any of these overlap with what you’re thinking about, let’s talk
02 · research workflow
Research operating systems
Reproducible experiment loops, agent-manageable repos, provenance, hypothesis trees, and tooling that helps research move without losing context.
03 · models & structure
Structured representations
How models encode position, geometry, and topology — from ViT positional encodings and RoPE to persistent homology and topology-aware representations.
04 · closed loops
Autonomous experimentation
End-to-end research loops for dry labs and wet labs: systems that design experiments, measure outcomes, update beliefs, and decide what to try next.
GPU scheduling
IaC
Slurm
ViT positional encoding
RoPE
Persistent homology
Generated-media detection
Explainable AI
Reproducible research
Self-driving labs
AI agents
Project traces
Things I keep building around
Infrastructure
LAIT compute & service stackclusters / monitoring / IaClab-scale operations
Making shared research compute easier to operate
Ansible, Slurm, GPU dashboards, service deployment, and admin workflows for a lab where compute is always the bottleneck.
Research systems
Agent-ready experiment workflowsagents / provenance / runsreproducible loops
Turning research state into something agents can use
Experiment control planes, source-of-truth repos, job routing, run lineage, paper ingestion, and hypothesis state that survives context switches.
Autoresearch
Closed-loop experimentsdry lab / wet labautonomous experimentation
From agent-run training loops to self-driving labs
Systems that propose experiments, run them, measure outcomes, and learn what to try next — whether the experiment is on a GPU or a bench.
Questions
Current questions
01
How should a small research lab allocate and operate GPU clusters without turning admin work into a full-time job?
02
What does an AI-manageable research repo need so agents can help without corrupting provenance or hiding failures?
03
Can positional encodings, RoPE variants, and topological signals make vision models understand structure better?
04
What is the smallest useful closed-loop experiment system — in software, dry lab, or wet lab — that teaches the right lessons?
05
How do we detect, explain, and evaluate AI-generated media as generation models keep improving?