Om Agrawal

Portfolio · 2026

I build things end-to-end and write about it.

CS at UT Austin. Researching inference systems with UTNS. Spent last summer shipping agent tooling at Anduril; joining Canyon Code this fall. Trained GPT-2 from scratch on 4 A100s; writing about the stack as I learn it.

Om Agrawal
Now Joining Canyon Code as an ML Systems Engineer (Fall 2026 Co-op); researching diffusion-model inference caching at UTNS. SF · Austin

Selected work

2024–2026 · 15 projects
01/ ACID: Adaptive Caching for Video Diffusion A training-free mechanism that switches feature-caching thresholds during denoising, keyed on the second derivative of the underlying method’s drift signal. Up to 2.16× over the no-caching baseline and 38% over conservative fixed-threshold caching, across 3 video diffusion models and 3 caching methods, with negligible quality loss. PyTorch · Video Diffusion · Inference Read the paper on arXiv → ACID paper, first page
02/ Lattice Agent Skills Agent evaluation infrastructure for Anduril’s Lattice APIs. The skills teach coding agents to integrate across languages, REST/gRPC protocols, and auth methods; an agent benchmark grades what they produce against live deployment state, and a CI feedback loop reruns it weekly to flag regressions as the APIs evolve. Agent Benchmarking · Evaluation Infrastructure · Feedback Loops · CI GitHub → Docs → Lattice SDK skills documentation
03/ LLM Stack from First Principles Worked through Karpathy’s Zero to Hero: autograd engine, MLP, batchnorm, WaveNet, char-level transformer, BPE, GPT-2. Wrote a technical blog covering each. PyTorch · Neural Nets · Transformers LLM Stack from First Principles 04/ Arbor A Git-like chat session manager for coding agents. Open-source MCP server + a modified mini-swe-agent for benchmarking. Python · MCP · SWE-Agent Arbor 05/ Cornucopia AI-powered fridge inventory tracker that reduces food waste and suggests recipes from CV-detected items. 1st place + Best Use of GenAI at IEEE UT’s 2025 Techathon. Python · Gemini · ESP32 · PostgreSQL Cornucopia

Recent writing

10 posts
Mar 20, 2026 Reproducing GPT-2 on 4 A100 GPUs read → Feb 21, 2026 Tokenization read → Feb 19, 2026 Building GPT read → Aug 28, 2025 Building a WaveNet-Inspired MLP Architecture read → Aug 18, 2025 Backprop on Tensors: Understanding PyTorch's Autograd Engine read →

Experience

Fall 2026
Canyon Code ML Systems Engineer Co-op INCOMING
Working on the agent dependency graph that batches and prioritizes model calls across multi-agent workflows.
Summer 2026
Anduril Software Engineering Intern
Shipped the official Lattice Agent Skills (docs), teaching coding agents to integrate against Anduril’s Lattice APIs. Built the agent evaluation infrastructure behind them: an agent benchmark graded against live deployment state, and a CI feedback loop that reruns it weekly to flag regressions.
2024–present
UT Networked Systems Research ML Researcher
Built ACID, a training-free mechanism that switches feature-caching thresholds mid-denoising, giving up to 2.16× faster video-diffusion inference with negligible quality loss. With postdoc Saurabh Agarwal and Prof. Aditya Akella.
Summer 2025
Boeing · ForeFlight Software Engineering Intern
Designed a unified networking protocol standardizing communication across 50+ ForeFlight-compatible avionics devices, plus a multithreaded Rust UDP client to validate the spec before vendors built against it.
2023–2025
Federal Aviation Administration Software Engineering Intern
Built a GPS-interference detection algorithm over 50M ADS-B datapoints, and a 3D statistical model localizing jamming/spoofing devices from 29M points in degraded-GPS scenarios. Presented findings to the FAA’s Chief Scientific Technical Advisor for Satellite Navigation Systems.
2023–2024
UT Austin CS Course Assistant, Discrete Mathematics
Taught discussion sections and led exam reviews for 150+ students; graded assignments and exams.

Education

Expected 2027
The University of Texas at Austin Integrated Master’s, Computer Science
B.S. + M.S. in Computer Science through UT’s Integrated Master’s Program.

Get in touch