Sarbjit Kaur

Consumer Attention Analytics: Eye Tracking

Clients needed to know exactly where consumers look on product packaging. Not what they say they looked at, but where their eyes actually went. We built a fully automated system: upload a webcam recording and get back attention heatmaps with zero human intervention.

A 3-container microservices architecture (Scheduler, Gaze Prediction, Calibration) processes video through a PyTorch gaze model, calibrates per person, and produces KDE heatmaps, fixation maps, and scanpaths. Deployed on AWS SageMaker. I optimized the calibration workflow to run 30% faster by cutting redundant head-pose computations.

Consumer Emotion Analytics: Video Analysis

Instead of asking people if they liked a product (which gives biased answers), we read their faces. Frame by frame, the system detects 7 emotions, 8 facial Action Units, and continuous valence and arousal scores. Researchers get objective data about genuine reactions.

Deployed via SageMaker async GPU endpoints with retry logic and heartbeat monitoring. Zero data loss across hundreds of sessions.

Food Trend Analytics Platform

What's the world eating? We scrape restaurant data across Indonesia, Thailand, Philippines, Malaysia, Vietnam, and India. The data is noisy and multilingual. I discovered the old regex cleaning was actually destroying valid information, so I redesigned it with LLMs (OpenAI Batch API) to translate, clean, and structure properly.

SentenceTransformer embeddings classify everything, and clustering surfaces seasonal trends by cuisine, dish, and eatery format. Multiple food companies now use this to decide what products to launch.

Multimodal AI Storytelling Platform

Kids speak a story idea, or draw it, or type it, and the system brings it to life as illustrated images with narration. Whisper handles multilingual speech input, Stable Diffusion generates the visuals, and fine-tuned YOLO makes objects interactive: tap an object, hear what it is.

The pretrained YOLO failed completely on animated, stylized imagery, so I directed a team of interns to curate our own training dataset from scratch and fine-tuned the model on it.

Agriculture Ecosystem Platform

A farmer in Punjab opens the app and gets: live sensor readings from their field, satellite imagery showing crop health (NDVI), ML-powered crop classification, weather advisories, and a real-time chat with agricultural experts.

I built the backend connecting all of this. Flask APIs, Celery for async ML inference, PostgreSQL for the data layer, Socket.IO for live communication.

IoT ECG Health Monitoring Device

Clinical-grade heart monitoring shouldn't require expensive hospital equipment. We built a 6-lead ECG device running TensorFlow on a Raspberry Pi: on-device ML inference for cardiac diagnosis with a real-time GUI so both doctors and patients can see the data.

Upgraded the system from 1-lead to 6-lead, handling multi-channel signal acquisition on embedded hardware. Published the research at ICCMR 2018.

LLM on Raspberry Pi

Deployed a Large Language Model on constrained hardware: a Raspberry Pi. Optimized for real-time inference speed and memory efficiency. Because why not.

Insurance Data Analytics

Automated the messy quarterly process of turning raw insurance claim files into standardized reporting datasets using Dataiku and SQL.