Aaryan Patil

United States0

Berkeley

Berkeley CS + Data Science student building distributed software systems for physical AI robots. Combines foundation model research with production systems shipping real multi-step task execution.

AI Fluency

No AI stories yet

Nothing posted yet.

Q&A

No questions yet

Nothing posted yet.

Projects

Building GPT from Scratch

Transformer implementation from first principles

Implemented a full transformer end-to-end from scratch: byte-pair tokenization, multi-head causal self-attention, positional embeddings, and the full training loop without high-level modeling libraries.

VisionScreen

Webcam-based vision screening battery

Built a browser-based vision screening battery with a dense encoder-decoder performing semantic segmentation of the eye. Trained locally on an M4 Max using PyTorch's Metal backend.

Mentora

AI voice tutor with spatial memory canvas

Built a voice-interactive tutoring system with a spatial memory canvas that speaks, listens, and renders visual components. Implemented multimodal RAG combining voice, visuals, and chat history.

Experience

Software Engineer

Walmart

Oct 2026 — Present

COO

Stealth Startup

Jan 2026 — Present

Undergraduate Machine Learning Researcher

Berkeley Artificial Intelligence Research

Feb 2026 — Present · Berkeley, CA

Third author on ICRA 2027 paper introducing a context-conditioned vision-language-action model that learns new tasks zero-shot from 1–2 demonstrations on a frozen foundation-model backbone. Implemented cross-attention conditioning and short-term memory modules in PyTorch; diagnosed and fixed a bf16 numerical overflow that reduced auxiliary loss from ~1.0 to ~0.009. Owns multi-GPU distributed training and architecture ablations on Berkeley's compute cluster and built automated benchmarking pipelines evaluating policies across simulation suites.

Co-Founder & Software Engineer

breefly

Aug 2025 — Mar 2026 · Berkeley, CA

Built and scaled a consumer web platform compressing long lectures into structured video. Grew to 7,000+ users and 2M+ organic impressions, raised a pre-seed round at $3M valuation. Owned the backend including transcript ingestion, contextual compression pipeline, asynchronous media-processing job queue, and the REST API serving the web client.

Education