Tadka

People describe what they want to cook in sentences, with constraints: "spicy without onions", "Jain food", "high protein, no dairy". Keyword search ignores those constraints, and chat answers are slow and come back as free text.

Tadka is a mobile-first app with 68 classic Indian dishes. Type a craving and, half a second after you stop typing, it ranks every dish against it: top picks as a swipeable card deck, the rest as a ranked list, and each dish as a recipe sheet you can save to an on-device cookbook.

Each search is one request to Jev, TypeSafe's decision model. Eleven questions read the craving itself (diet, spice, meal, how heavy, time, region, and which ingredients to avoid), and 68 yes/no questions ask whether each dish fits.

Plain code turns those answers into the order: Jev's fit, multiplied by diet, preference and avoid factors. Diet acts as a near-hard rule, preferences only nudge, and an ingredient is penalised only once Jev is at least 30% sure you want to avoid it. Every dish shows the factors behind its match, and when nothing fits every part of a request the list says "Closest options" instead of pretending.

Tadka is a progressive web app: Chrome reports it installable with no errors, and a service worker caches the app shell and dish photos. Offline, the installed app still opens and search falls back to a keyword matcher that says so on screen. Dark mode follows the phone's setting, and identical searches are served from a cache in milliseconds.

Across 17 fresh searches, Jev answered all 79 questions in a median 251 ms (123 to 352 ms) using about 10,350 input tokens, about $0.00044 per search. "Light vegetarian dinner" returned dal tadka, bhindi masala and khichdi; "something sweet for Diwali" returned gulab jamun, gajar ka halwa and kheer; and "pizza" sensibly came back as closest options only.