Designed a transformer-based actor-critic architecture with a novel tokenization scheme to train a single reinforcement-learning policy across multiple robot morphologies simultaneously. Conditioned the policy on BERT-encoded URDF (robot-structure) descriptions, enabling generalization to robot embodiments unseen during training, without per-robot retraining.
Optimized a semantic-segmentation model for real-time scene understanding on an autonomous-vehicle platform, targeting the NVIDIA Jetson ORIN edge device. Converted the PyTorch model to an INT8-quantized TensorRT engine (calibrated on ~6,000 images), achieving a 3× inference speedup (from ~50 to ~150 fps) while maintaining an acceptable accuracy threshold.
Developed and implemented an actor-critic algorithm to solve constrained optimization for cooperative multi-agent reinforcement learning. Evaluated the approach on two benchmark multi-agent environments — Cooperative Navigation and Cooperative Treasure Collection — using an Azure cloud virtual machine. This work led to a publication at AAMAS 2021.
Implemented research papers for customer-specific use cases across reinforcement learning, NLP and computer vision. Selected projects include Explainable Product Recommendation, Airline Seat Inventory and Dynamic Pricing, and Fine-Grained Visual Classification.