ML | Builder & Leader | CS @ UIUC
Chicago, IL | rheats2@illinois.edu | linkedin.com/in/rheatshah | rhea-shah23.github.io
Marrow: A dependency-free LLM inference engine for Apple Silicon that writes its own ARM64 kernels.
- C11, zero dependencies — runs Llama models with nothing but libc.
- Writes its own machine code — hot kernels are compiled and tuned at load time.
- Measured against the hardware limit, not against nothing — speed as a fraction of the machine's real memory bandwidth.
TweetSim: Predict Your Tweets Engagement BEFORE You Post!
- Leverages xAI's opensource X feed algorithm to score users' tweets
- Suggests improvements to encourage engagement
ARGreeksLab: iOS + ARKit App Development
- Visualizes Black-Scholes as something you can walk around
- Ideally an internal tool meant to develop intuition
- Turns options prices and greeks into interactive 3D surfaces in augmented reality
- Startup
- Dashboard of user's messages with actionable items, ideas, and any other important information
- Context-aware notification: location-based, time-based, person-based
2024 NeurIPS Spotlight Project: Advancing Diabetic Retinopathy Diagnosis Using Vision Transformers
- Engineered a Vision Transformer in PyTorchLightning for 5-class diabetic retinopathy diagnosis using the RetinaMNIST dataset
- Deployed patch embeddings, self-attention, and positional encodings to capture spatial-retinal patterns with high accuracy
- Surpassed all published benchmarks, including Google AutoML Vision, setting a new state-of-the-art benchmark
- Built a real-time risk engine in Rust to compute PnL, exposure, and liquidation metrics for high-frequency crypto derivatives trading
- Integrated Kafka and TimescaleDB to support millisecond latency data ingestion and scalable time-series risk aggregation
- Deployed Grafana dashboards to monitor system-wide financial health, enabling sub-second insights for automated trading decisions
- Partnered with MIT Lincoln Laboratories
- Engineered a deep learning pipeline to classify 3D printing failures using a curated 10K-image subset of the CAXTON dataset
- Benchmarked ResNet, ViT, and Swin models with ResNet50 achieving 89% accuracy
- Used feature extraction to cut model size by 99.9%
- Resolved data quality issues and conducted model analysis to inform autonomous defect detection in additive manufacturing
I'm always open to collaborations, internships, and new research opportunities!
Phone: 815-616-7848
Email: rheats2@illinois.edu
LinkedIn: Connect with Me!
Location: Champaign, IL (Open to Remote/Hybrid AND In-Person Roles)