Manjunath Mohan

Leading the development of AI models for robotic cooking automation. Built and deployed a MobileViT model for cooking stage detection with 0.98 F1 Score, achieving 15 FPS real-time inference on ARM Cortex CPUs. Designed and scaled Data Collection and human-in-the-loop labeling system for ingredient and cooking-stage datasets. Developed and deployed ResNet and MobileNet models to classify and track ingredient stage cooking progression with 0.99 F1 score. Implemented an optimized C/C++ inference module that detects NPU availability and automatically dispatches models to it with fallback to CPU with ONNX and TFLite backends. Interfaced with Android. Built an LLM-based transformation engine that converts and adapts arbitrary online recipes into Nosh executable cooking program.

Part of the analytics development team, focusing on computer vision applications for CCTV surveillance. Designed and implemented a dual-Camera Calibration PoC for reliable cross-camera object tracking across city-scale CCTV networks. Developed and deployed real-time Fire, Smoke Detection System using surveillance feeds in a Petroleum Refinery. Developed cameraworld geometry-based logic for abandoned baggage detection using extrinsic calibration to track object permanence and owner separation in real-world space. Implemented a novel false-positive filtering algorithm for fire, smoke detection, reducing 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. Designed and deployed a deep learning-based Eye / Gaze Tracking system using a single NIR camera, achieving 3 RMSE within an automotive Driver Monitoring Suite. Architected and scaled a high-precision Data Collection and Automated Ground-Truthing system for Gaze, Eye Position, Head Pose, and Driver Position using a Multi-Camera Motion Capture Setup and Kinect. Enabled parallel multicamera (PoV) capture to accelerate dataset acquisition. Contributed and later led core perception modules for an active driver distraction system, including face detection, head pose estimation, and pupil tracking. Improved runtime performance through structured Pruning and Quantization (including Quantize-Aware Training), enabling combined DMS perception stack to run at 10 FPS on low-powered automotive hardware without accuracy loss. Developed a comprehensive evaluation suite aligned with data collection workflows to benchmark and validate DMS performance. Fine-tuned and validated perception modules across multiple OEM-specific camera placements and configurations.