INCOMING AT UCLA · ROBOTICS SOFTWARE LEAD · AI/ML RESEARCH
RIAN
BUTALA
I reimplement modern ML from the papers up: diffusion, JEPA world models, and LLMs built from scratch in PyTorch. Then I build the perception systems and apps that put them to work.
01 · FOCUS
01
Deep learning, from scratch
Diffusion (rectified-flow latent), JEPA world models, and LLMs reimplemented from the papers up in PyTorch.
02
Vision & perception systems
Facial recognition, autonomous surveillance, and VLM-driven captioning and anomaly analysis over live video.
03
Full-stack & tooling
Health and productivity apps, agent platforms, and developer tooling, down to a GPU monitor for macOS PyTorch.
STACK
ML / RESEARCHPyTorchCUDADiffusionJEPAQuantization / QATComputer visionYOLOVLMs
LANGUAGESPythonC++JavaTypeScript
SYSTEMS & WEBFastAPIFlaskNext.jsArduPilot
02 · SELECTED WORK
// from github
Latent diffusion from scratch
Rectified-flow latent diffusion in an MMDiT-style architecture, written from the paper up. The fifth iteration over eight months: earlier attempts wouldn't train stably until I rebuilt the data pipeline. The live demo samples 256px images from my trained checkpoint.
▶ LIVE DEMO ↗PyTorch
Homegrown LLM
A transformer language model pretrained from scratch, then instruction-tuned. The third in an eight-month line of from-scratch LMs, after an autocomplete model and a first chat model. Chat with it in the live demo.
▶ LIVE DEMO ↗PyTorch
OpenAI Parameter Golf
OpenAI's challenge to train the best language model under 16 MB. Mine reached 1.5126 bits/byte in a 600-second window on a single A5000.
PyTorch
V-JEPA world model
A self-supervised video world model (V-JEPA) hand-written from scratch in PyTorch. Training was stable, no representation collapse and loss converging, though I did not run downstream evaluation.
PyTorch
OtterAgent ★ BEST AI HACK
A social network where AI agents negotiate on your behalf, managing calendars, notes, and tasks under scoped permissions. Won Best AI Hack at Synthesis Hacks @ Google.
DEVPOST ↗FastAPI · Gemini
Vesta, surveillance engine
Built for Stratford Preparatory to deter and detect theft across multiple cameras. YOLO triggers recording on people; a vision-language model scores each clip for threats and makes them searchable in plain language. Runs entirely on local hardware, no cloud.
Flask · YOLO
btop, with macOS GPU
GPU telemetry in btop for Apple-Silicon PyTorch users, no sudo required.
C++
03 · ABOUT
I optimize for efficiency: getting models to train fast, fit in tight budgets, and run on the hardware I have. Most of my work is reimplementing recent papers from scratch to understand them end to end, then building real systems on top.
04 · EXPERIENCE
Research Assistant · UCLA
Invited as a high school senior to analyze neural recordings and optimize decision transformers with graduate students and faculty.
2026 - now
Software Engineer · WakeMates
Sole developer taking the product from concept to deployment alongside the founder.
2026 - now
Founding Software Lead & Captain · FRC 10252 / FTC 21781
Built a rookie team's full codebase from scratch and mentored members in computer vision and autonomous path planning.
2022 - 2026
Founding VP, Aerospace Club · Stratford Preparatory
Built and flew autonomous long-range RC aircraft from scratch. Pixhawk, ArduPilot, ExpressLRS.
2025 - 2026
AI/ML Intern · FinanceGPT
Machine learning research and applied work over a summer internship.
2024
05 · EDUCATION
UCLA
B.S. Computer Science.
2026 - 2030
Stratford Preparatory
National Merit Finalist. Robotics, Aerospace Engineering, Tennis.
2022 - 2026
06 · GET IN TOUCH
Open to internship and research opportunities. The fastest way to reach me is email.
rianbutala@gmail.com