There is a lot of information in your blood and bone marrow cells. At the Heme AI lab, we develop deep learning models and diagnostic tools to help pathologists improve their workflow and get insights about a patients cancer.Key contributions:- Developed and productionized “CAREMIL”, a flagship model of our lab that achieves SOTA performance on whole slide images for hematological tasks. Submitted to NeurIPS.- Helped recruit and manage new members to expand our research into new projects- Conduct user interviews to build product and business strategy to spinoff the lab into a companyReach out to know more.

Developed a human in the loop RL method that predicts the variance of returns and flags high risk decisions. Applications include robot navigation and manipulation. Published our work to ICRA 2024

Supply Chain Optimization TechnologiesWorked in the selection team which chooses which products must be delivered to customers on the same day. Designed a customer choice model that uses a novel logistic classifier to predict the utility of every product and the purchase probability for a product group. Improved the purchase probability prediction by 38% across all product groups.Implemented the training, inference, and evaluation of this ML model in python and pushed my code changes to production.Presented my findings to the entire team and wrote a research paper that will be used in internal ML conferences.

Worked in a team of 4 to simulate robotic grippers in MuJoCo with force-controlled actuators for manipulation tasks.Collaborated with university researchers to design and test reinforcement learning based systems for the simulator using Keras to make contact rich robotic grippers.

Built planning and control algorithms for autonomous vehicles using CARLA simulator in an urban environment.Experimented with modified planning algorithms to use RL for waypoint speed computation using DQN and DDQN. The results gave the company a new direction to train their vehicles.Proposed inverse RL and imitation learning methods as an alternate which reduced tunable hyperparameters while improving driving performance.Worked under the supervision of the CEO to program GAIL, PPO and actor critic methods using PyTorch to test and compare their performance in CARLA and presented my findings to the entire team.

Designed and deployed an object detection system for a UAV to identify people and vehicles from an 80m height.Cleaned 1000+ GB of annotated image data from various sources to create a trustable repository for the ML team.Trained and finetuned a TensorFlow single shot detector on the custom dataset to achieve 92% object detection accuracy.Compressed the model by 30% and tested it on the drone using raspberry pi to get 14fps during inference.

Worked on real time lane finding, semantic segmentation, and PID Control for autonomous vehicles.Trained and deployed a U-Net segmentation in Tensorflow with 62% IoU accuracy to assist autonomous truck navigationUtilized a stereo camera system to predict depth in an image which improved the information from the vehicle sensors.