Manjunath Mohan

Leading the development of AI models for robotic cooking automation. Built and deployed a CLIP + Vision Transformer model for ingredient-stage detection, achieving 8 FPS real-time throughput on ARM Cortex CPUs. Designed a data collection and human-in-the-loop active learning pipeline for ingredient and cooking related data curation. Developed multi-stage classification models based on ResNets and ViTs to monitor ingredient cooking progress. Designed and implemented optimized C modules for inferencing models on ARM chips with ONNX and TFLite backends with GPU/NPU passthroughs and interfaced with Android. Created an LLM-powered recipe conversion app, transforming online recipes into machine-readable, executable cooking workflows for Nosh Cooking Robot.

Part of the analytics development team, focusing on computer vision applications for CCTV surveillance. Designed a dual-camera calibration system enabling robust object tracking across city-wide CCTV grids. Built and deployed fire-smoke and abandoned baggage detection systems using surveillance camera feeds for a petroleum refinery. Implemented novel false-positive filtering algorithm for fire and baggage detection, cutting unnecessary operator reviews by 70%.

Core member of the algorithms team, specializing in ADAS features using computer vision and deep learning. Led DMS dataset acquisition efforts. Developed a high-precision eye-tracking system (sub 3 deg error) using a single IR camera, deployed for driver distraction monitoring. Contributed to multiple core driver monitoring modules (DMS), including face detection, head pose estimation, and pupil tracking. Designed and deployed a real-time data collection and ground truthing system for DMS integrating multiple cameras and depth sensors. Optimized deep learning models for low-power automotive hardware devices using pruning and quantization techniques. Built a Key Performance Indicator (KPI) calculator to assess and benchmark DMS performance across hardwares. Adapted and validated DMS perception modules for multiple OEM-specific camera placements.