Founded and engineered a Mongolian-first multimodal AI nutrition platform that combines a fine-tuned vision-language model, localized food data, and schema-validated outputs to turn meal photos into calorie and macro logs. The product reached 2,500+ users through repeated TestFlight releases, local payment work, and angel-funded grant support.
Founder & CTO
Global calorie trackers are built around English interfaces, Western food databases, and generic meal assumptions. Mongolian users needed nutrition tracking that understood local language, local dishes, mixed plates, and low-friction logging behavior.
A React Native mobile product with manual food search, USDA nutrition data, translated ingredient metadata, and AI photo logging routed through Supabase Edge Functions. The vision pipeline used a fine-tuned VLM trained on 400 labeled Mongolian food images, then converted inference into schema-validated calories and macronutrients, reconciled results against local-food records, and persisted the final entry into the user’s daily nutrition state. I also owned onboarding, subscriptions, analytics, TestFlight release loops, and coordination around local payment and bank APIs.
Reached 2,500+ users, built a 300+ person waitlist before launch, ran 8+ TestFlight iterations, collected roughly 400 local food images, onboarded 3–4 interns, and received an angel-funded grant to continue building.
Localization is deeper than translation: the model distribution, food ontology, fallback search, database, onboarding, payment flow, and product language all have to fit the community’s real behavior. Multimodal AI only becomes a product when uncertain recognition has safe fallbacks and the output is deterministic enough for downstream calculations.
The constraints, reliability decisions, and system boundaries behind the visible product.
Treated localization as a systems problem: language, food data, image recognition, onboarding, payments, and nutrition output all had to fit Mongolian users.
Routed AI photo logging through Supabase Edge Functions and structured calorie/macro output so the mobile app received predictable data instead of raw model prose.
Improved the vision loop with local food images, manual fallback search, USDA data, translation, and local database matching rather than depending on one model call.
Built the multimodal boundary around a fine-tuned VLM and schema-validated nutrition objects, then reconciled uncertain predictions against local food records before committing calories and macros to user state.
Owned the production surface around the model—Edge Function inference, authentication, subscriptions, analytics, TestFlight distribution, and local payment coordination—so recognition quality connected to a usable consumer product.