
Research Engineer
Novus Hi-Tech Robotics Systemz · Jul 2023 – Present
- Developed and deployed production computer-vision models for an ADAS driver-monitoring stack covering inattention, drowsiness, seatbelt, and smoking detection, from model training through on-device inference.
- Designed and developed a facial landmark detection model, expanding facial landmark prediction from 6 to 68 points; fine-tuned it on Indian-face data and deployed the model using TFLite.
- Reduced facial landmark model inference latency from 50 ms to 30 ms (~40%) on the target edge device through model and inference optimization.
- Developed a lightweight ~350K-parameter gaze network combining left/right eye crops, face CNN features, and graph-attention fusion; trained on ETH-XGaze, RT-GENE, and MPIIFaceGaze, achieving ~4-5° mean angular error.
- Managed ML experimentation and release workflows using MLflow, Docker, CI/CD, and Git, with model validation between training/Keras outputs and deployed board inference.




