I'm Pranav Somu

Led a team of 16 to engineer a novel surrogate for predicting model performance, using LLMs to understand the models’ code (published GECCO 2026), achieving performance competitive with prior research using 97% less training data.

Developed a GenAI physiotherapy trainer as a Swift app, automating personalized treatment planning using a LangChain RAG pipeline with GPT-4o. A deployed text-to-motion model converts these plans into animated GIFs, offering accessible and clear visual guidance for exercises.

Directed a team of 18 in creating a service to support data-driven agricultural policy by forecasting crop type and yield using satellite data. Led the optimization of LSTM and UNet models to achieve 91%+ prediction accuracy.

Developed a supply distribution system for a non-profit, enabling the efficient completion of over 800 yearly requests for 300 foster families. Improved system performance and data query speeds by 220x by refining the Python backend, Django REST APIs, and PostgreSQL database.

Engineered deep learning algorithms for accessible ASL education through a real-time, ASL-to-English translation system. The solution uses a 95% accurate SSD ResNet50 hand sign classifier and a transformer architecture that surpassed prior research accuracy by 2%.

Developed an intelligent restaurant search and recommender engine featuring simple filtering and demand analytics for businesses. Fine-tuned a BERT model to classify attributes like cuisine from plaintext searches, achieving 93% accuracy in understanding user intent.