Manjunath Mohan

Leading the AI team in building culinary AI Systems to automate cooking. Ingredient Understanding - Built and fine-tuned ResNet, MobileNet, MobileViT and Swin Transformer models to infer shape, size, variants, and cooking state across many ingredients, with F1 > 0.95 on diverse test data for automated cooking decisions. Ingredient Segmentation - Fine-tuned SegFormer to segment different ingredients in the pan with mIoU of 0.94, enabling volume estimation to drive cooking logic. Universal Consistency Model - Developed a MobileViT model to classify dish consistency progression (liquid-to-dry) across 4 stages for adaptive cook-time and heat controls. Data Engine - Led data collection and human-in-the-loop labeling efforts to build high-quality culinary datasets. Vision Monitoring - Built a MobileViT model to detect camera degradation factors (steam blur, oil splatter, exposure shifts) and early sticking and burn detection, improving downstream model reliability and cooking outcomes. Edge AI Deployment - Built an optimized C++ inference system with dynamic NPU/GPU/CPU dispatch, supporting ONNX and LiteRT runtimes, deployed on Android. Recipe Intelligence - Built an LLM pipeline converting free-form online recipes into structured cooking programs, with modality adaptation (gas to induction, closed to open pot), achieving a 40% A/B preference rate compared to chef-curated machine tuned recipes.

Took a career break to address a critical family medical situation while overseeing family-owned business (housing construction projects and related equipment rental ops).

Member of the analytics development team, focusing on computer vision applications for CCTV surveillance. Fire and Smoke Detection System - Developed and deployed real-time fire and smoke detection system on petroleum refinery surveillance feeds, reducing manual operational review through false-positive filtering by >75%. Abandoned Baggage Detection System - Built abandoned-baggage detection using camera-world geometry, extrinsic calibration, object permanence, and owner-separation logic for CCTV analytics. Multi-Camera Tracking PoC - Designed a dual-camera calibration proof of concept for cross-camera object tracking across city-scale CCTV networks.

Member of the DMS algorithm team, specializing in ADAS features using computer vision and deep learning. Led DMS dataset acquisition efforts. Eye / Gaze Tracking - Designed and deployed a single-NIR-camera eye/gaze tracking model for an automotive Driver Monitoring System, achieving 3 RMSE. Data Pipeline - Built a high-precision data acquisition and automated ground-truthing pipeline for gaze, eye position, head pose, and driver position using multi-camera motion capture and Kinect. Driver Distraction System - Contributed and later led core perception modules for driver-distraction detection, including face detection, head-pose estimation, and pupil tracking across OEM-specific camera placements. Model Optimization - Optimized the DMS perception stack with pruning and quantization-aware training, enabling 10 FPS on low-power automotive hardware without accuracy loss. DMS Evaluation Framework - Developed a comprehensive evaluation suite aligned with data collection efforts to benchmark and validate performance across DMS providers in diverse scenarios.