Indian Food Recognition & Automated Logging
Machine Learning Lead @ HealthifyMe — May 2018 to Jul 2019
Overview
Faced with high dropout rates due to manual food logging, we built India’s first AI-powered “snap‐and‐log” system, proving feasibility to our CEO/CTO in less than 2 weeks. Over 1 million Indian‐food images were collected to train a bespoke deep‐learning classifier, achieving > 83 % accuracy on diverse regional dishes. Thousands of users now log meals via a single tap.
Work
- Rapid Prototype (2 weeks):
- Curated an initial dataset of 5 K+ home‐cooked dishes.
- Trained a lightweight CNN from scratch; deployed behind a simple Flask API and basic web UI.
- Scale & Accuracy:
- Expanded to a 1 million–image Indian Food Dataset covering 1 500+ dish classes.
- Optimized model for < 100 ms inference on CPU; integrated quantization and pruning.
- Production Deployment:
- Built a production service for real‐time image upload, inference, and calorie estimation.
- Instrumented analytics to track usage and guide iterative improvements.
Impact
- User Engagement: Snap‐and‐log feature increased meal‐logging retention by 60 %.
- Business Value: Convinced leadership to spin out a dedicated Computer Vision team.
Resources