Inspiration
Nigeria wastes an estimated 40 million tonnes of food annually while 25 million people go hungry. We wanted to solve this locally, not globally. Under Nigeria's National Home-Grown School Feeding Programme (NHGSFP), school cooks prepare meals for thousands of students daily — but quantities are based on total enrollment, not who actually shows up. On rainy days, exam weeks, or festive periods, hundreds of portions go uneaten. At Ataoja School of Science in Osogbo, Osun State, that means a cook preparing for 3,000 students when only 2,200 are present — every single day. We asked: what if the cook had an AI that already knew today's attendance, today's weather, and how much students liked today's dish — before a single pot was put on the fire?
What it does
WasteRadar NG is a native Android app that helps the school cook at Ataoja School of Science predict exactly how much to prepare each day, in local Nigerian units (kongo, bags, sachets, tubers), priced at real current Osogbo market rates. Every morning, the Admin enters class-by-class attendance (JSS1–SS3). The AI Reasoning Engine — powered by Groq's openai/gpt-oss-120b with reasoning_effort: high — then weighs the confirmed attendance against today's dish preference score, rain intensity, school calendar type, and supply status to recommend the exact quantity of that dish's main ingredient — for example, 2 bags + 8 kongo of rice, or 200 sachets of spaghetti — in the local units Osogbo market vendors actually sell in, alongside a confidence percentage and plain-language explanation. When surplus is predicted, the app performs a live web search near the school's GPS coordinate (7.7485979°N, 4.5170886°E) and surfaces two real nearby redistribution locations — orphanages, food banks, or NGOs in Osogbo — with embedded maps and walking distances. The app also shows the full ingredient breakdown in local units with real Naira costs, calculates money saved versus cooking for full enrollment, and learns weekly from actual logged consumption, updating each dish's preference score automatically. The cook reviews and approves every recommendation before cooking begins. The AI never acts alone.
How we built it
Android (Kotlin + Jetpack Compose + Material 3) — native app, MVVM architecture, Room database for local persistence Groq API (openai/gpt-oss-120b, reasoning_effort: high) — the core AI reasoning engine; pattern detection, predictive modeling, and recommendation generation Groq browser_search tool — real-time search for nearby redistribution locations near the school coordinate Firebase Remote Config — secure API key storage without hardcoding in the APK Knowledge Base JSON — a structured domain file encoding Nigerian school feeding patterns, NHGSFP context, dish preference scores, and local unit conversion constants, injected into every AI call to ground the model's reasoning in Osogbo reality Google Maps free embed — zero-billing map previews for redistribution locations AI Studio — used to generate and iterate the app codebase We ran six AI research tools (ChatGPT, Perplexity, Gemini, Kimi, DeepSeek, GLM) to ground every design decision in real Nigerian food waste data before writing a single line of code.
Challenges we ran into
Keeping AI reasoning genuine — the biggest challenge was ensuring the model was actually deciding the final portion count, not just narrating a number computed by a rule engine. We built the rule engine as a baseline reference only, and the model holistically weighs all factors — including conflicting signals like heavy rain on the most popular dish day — before arriving at its own recommendation. Local unit conversion — Nigerian cooks don't think in grams. Building a Kongo/bag/sachet/tuber conversion engine that matched how ingredients are actually bought in Osogbo markets took significant research. API key security on Android — APKs can be decompiled. We avoided hardcoding the Groq key by routing through Firebase Remote Config, with a plan to rotate the key post-submission. Attendance as ground truth — replacing predicted attendance with real class-by-class data meant redesigning the AI's role entirely. Rain and calendar type no longer predict whether students show up — they now model how many present students rely on the school meal versus alternatives like outside vendors or packed lunches. We also caught and fixed a subtle bug where the model's free-text explanations occasionally invented their own ingredient numbers, conflicting with the app's own calculated figures. The fix was structural: the model is now constrained to provide only causal reasoning in plain language, while all ingredient quantities are calculated once, by the app's own code, and never recalculated by the AI in prose. This reinforced an important lesson — when an AI explains a number, it should never be allowed to silently recompute that number differently.
Accomplishments that we're proud of
A fully working native Android APK targeting a real, named school in Osogbo with real GPS coordinates and real local market prices An AI that visibly reasons — the deep reasoning trace shows the model's actual thinking process, not a canned output A self-learning preference scoring system that updates weekly from real consumption logs, making the app genuinely smarter over time Real-time redistribution location search using Groq's browser tool — no paid Maps API, no hardcoded NGO names Full Nigerian Pidgin English toggle across all cook-facing screens, because a busy canteen cook in Osogbo should use a tool that speaks her language
What we learned
A spreadsheet can store data. AI earns its place when signals conflict — when it's raining heavily but today is the most popular dish day, when exam week overlaps with a festive period, when supply arrived late but attendance is high. Those are the moments where a rule-based system guesses and an AI reasons. We learned to design specifically for those moments, not the easy ones. We also learned that local specificity is not decoration — the difference between "a tool for schools" and "a tool for the canteen cook at Ataoja School of Science on a rainy Thursday in Osun State" is the difference between a generic demo and a real solution. The entire app — every prompt, every fix, every screen — was built and iterated on a mobile phone, with no laptop or traditional development setup. We learned that AI-assisted tools have lowered the barrier enough that a real, working native Android app is buildable from a phone alone, which matters in a context like ours where that's often the only device available.
What's next for WasteRadar NG
Expanding to other NHGSFP schools across Osun State A headteacher dashboard showing waste reduction trends across the whole school week Integration with local food rescue NGOs for automated redistribution alerts via WhatsApp Offline-first AI using a lightweight on-device model for schools with no internet connectivity
Log in or sign up for Devpost to join the conversation.