I build the agent core at DeepInvent — give it an arXiv ID and it reproduces the paper, then grades itself on what actually ran on disk, not on what the model claims.
CS + Math at UT Austin. Austin, TX.
CS + Math at UT Austin, class of 2028, in Austin.
I like problems where the failure modes are the interesting part — a scheduler that has to survive CUDA out-of-memory and a cloud running out of GPUs, a job pipeline with 18 states and five ways to recover, a research result that falls apart once you run the statistics properly.
Texas-raised, Nepal-rooted.
Reproducing ML papers from an arXiv ID.
Give it an arXiv ID and it reproduces the paper end to end. I built the agent core: 19 primitives the model calls to do the actual work, with the paper held as a REPL variable it never sees in its context — so it has to run things, not pattern-match the abstract. A result only passes when the measured on-disk evidence backs it; the model's own grade doesn't count. The other half is the GPU layer — one sandbox over local machines, Docker, GCP, and Azure/AKS, with a scheduler that pins each job to a single GPU, drops the batch size when CUDA runs out of memory, and moves a job from GCP to Azure when GCP has no capacity left.
Software Engineer · May 2026 – present · Remote
I own the agent core and the multi-cloud GPU layer.
Both retired — DeepInvent is where I build now.
Competition projects — all archived.
Built the agent core for reproducing ML papers from an arXiv ID — 19 model-callable primitives, results decided by measured on-disk evidence rather than the model's grade — and the multi-cloud GPU layer across local, Docker, GCP, and Azure/AKS: one GPU per job, batch-shrink on CUDA OOM, and GCP→Azure failover when capacity runs out.
May 2026 – present
One of two engineers taking the app from zero to 5,000+ users and 500+ App Store reviews at 4.7 stars. Owned the SwiftUI launch flows and Firebase sync, and built three StoreKit 2 subscription tiers with server-side receipt validation.
Jul – Oct 2025
Showed that DDS middleware saturation, not CPU, is what breaks a ROS 2 arm: 4,000 msg/s of flooding drove MoveIt on a simulated 7-DOF Panda to 100% execution failure, traced with LTTng and flame graphs. Replaying the same load under cgroups v2 on 32-core CloudLab nodes ruled compute out. Automated the 10-phase CloudLab setup so an experiment runs in minutes instead of hours.
Dec 2024 – May 2026
Threw out 8 attention-head findings on a 96-class Time Series Transformer that didn't survive 10,000-resample bootstrap CIs and Benjamini–Hochberg correction. Stress-tested 24 head rankings with Cohen's d under noise, time-warp, and phase-shift — steady on simple tasks (Spearman rho 0.88), falling apart on hard ones (0.48).
Jan – Feb 2026