Building AI for Indian palates

India has 28 states, 22 official languages, and a menu vocabulary that changes every 100 kilometres. Building AI that recommends the right dish to the right guest is not a copy paste of what Silicon Valley shipped.

Here is how we think about it.

1. Data grounded in Indian menus. Our menu understanding model was trained on real Indian menus from partner restaurants across Pune, Hyderabad and Bengaluru. It knows that paneer tikka and paneer 65 are not the same dish and that a Jain preparation excludes root vegetables.

2. Dietary reasoning first. Before recommending anything, the model checks the guest's dietary tags. Jain, vegan, diabetic, keto, halal, gluten free, nut allergy and low sodium are hard constraints, not suggestions.

3. Occasion context. A birthday dinner recommendation is different from a Tuesday solo lunch. The model factors in group size, occasion and time of day.

4. Restaurant intent. Chef's Picks, high margin dishes and celebration add ons get boosted when they match the guest's profile. This is how we lift AoV by 30 to 50 percent without pushing junk.

5. Feedback loops. Every thumbs up, every uneaten dish, every re order teaches the model. Restaurants see the aggregate insight in the merchant dashboard.

What is next. Multilingual voice ordering, regional palate models for 20 Indian cities, and Selfie Mirror pose recognition for personalized greetings.

We are hiring AI engineers and ML researchers. If teaching a model the difference between misal pav and pav bhaji sounds fun, come build with us.