Om Agrawal
Selected work
01/
One Threshold Is Not Enough: Prompt-Invariant Caching Schedules for Video Diffusion
One fixed caching threshold is not enough: a few critical denoising steps need conservative caching while the rest can be cached aggressively, and those steps are consistent across prompts. ACID is a training-free schedule calibrated once offline that pushes the speed–quality frontier across 3 models and 3 caching methods.
Read the paper on arXiv →
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.
GitHub →
Docs →
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.
04/
Arbor
A Git-like chat session manager for coding agents. Open-source MCP server + a modified mini-swe-agent for benchmarking.
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.
06/
Reproducing GPT-2 on 4× A100s
Full reproduction of 124M GPT-2 from scratch: bfloat16, torch.compile, FlashAttention, DDP. Trained on FineWeb-Edu 10B; beats the reference eval score on HellaSwag.
07/
Lock-Free Web Server
Multithreaded lock-free web server in Rust.
08/
Barnes-Hut N-Body
Efficient N-body simulation using the Barnes-Hut algorithm.
09/
2-Phase Commit in Rust
A 2-phase commit protocol simulation in Rust.
10/
K-means on GPU
GPU-accelerated K-means clustering algorithm.
11/
BST Lightning
High-performance multithreaded algorithm for identifying equivalent binary search trees.
12/
DDP Training
Distributed data-parallel training of VGG11 on CIFAR-10 with AWS SageMaker, with a scaling analysis of parallelism vs. gradient-sync overhead.
13/
LoRA Fine-Tuning
Parameter-efficient fine-tuning of TinyLlama with LoRA adapters in MLP layers, measuring training latency vs. rank and module size.
14/
Activation-Aware Quantization
Quantized OPT-1.3B to 3 bits while protecting the top 1% of salient weights via hardware-friendly scaling.
15/
Bhagavad-GPT
A RAG chatbot grounded in the Bhagavad Gita, referencing an authoritative knowledge base.
08/
Barnes-Hut N-Body
Efficient N-body simulation using the Barnes-Hut algorithm.
09/
2-Phase Commit in Rust
A 2-phase commit protocol simulation in Rust.
10/
K-means on GPU
GPU-accelerated K-means clustering algorithm.
11/
BST Lightning
High-performance multithreaded algorithm for identifying equivalent binary search trees.
12/
DDP Training
Distributed data-parallel training of VGG11 on CIFAR-10 with AWS SageMaker, with a scaling analysis of parallelism vs. gradient-sync overhead.
13/
LoRA Fine-Tuning
Parameter-efficient fine-tuning of TinyLlama with LoRA adapters in MLP layers, measuring training latency vs. rank and module size.
14/
Activation-Aware Quantization
Quantized OPT-1.3B to 3 bits while protecting the top 1% of salient weights via hardware-friendly scaling.
15/
Bhagavad-GPT
A RAG chatbot grounded in the Bhagavad Gita, referencing an authoritative knowledge base.
Recent writing
Experience
Canyon Code
ML Systems Engineer Co-op
NOW
Extending Nalar, the agent-serving runtime behind CanyonOS, to serve many agentic workflows at once on shared GPUs.
Anduril
Software Engineering Intern
Shipped the official Lattice Agent Skills (docs), teaching coding agents to integrate against Anduril’s Lattice APIs. Built the agent benchmark and CI feedback loop that keep them accurate as the APIs evolve, working with engineers at OpenAI, Anthropic, and Fern.
UT Networked Systems Research
ML Researcher
Built ACID, a training-free caching schedule that makes video-diffusion inference faster with negligible quality loss. With Saurabh Agarwal and Prof. Aditya Akella.
Boeing · ForeFlight
Software Engineering Intern
Designed ForeFlight Connect’s unified networking protocol for avionics devices, and a Rust client to validate it before vendors built against it.
Federal Aviation Administration
Software Engineering Intern
Built models to detect GPS interference and localize jamming/spoofing devices from ADS-B flight data; presented findings to the FAA’s Chief Scientific Technical Advisor for Satellite Navigation.
UT Austin CS
Course Assistant, Discrete Mathematics
Taught discussion sections and led exam reviews.
Education
The University of Texas at Austin
Integrated Master’s, Computer Science
B.S. + M.S. through UT’s Integrated Master’s Program. University & College Honors; IEEE UT Techathon 2025 winner.