EcoSnap Global

AI-powered waste intelligence that works online, offline, and via SMS — built for Africa, scalable to the world.

Comment

Inspiration

The trigger wasn't a statistic. It was a walk through a Neighbourhood in Lagos and watching a woman burn a pile of sachet water nylons on the roadside — not because she didn't care about the environment, but because she genuinely had no idea what else to do with them. No recycler had come. No bin existed. No app had ever reached her. Most climate-tech solutions are built for people who already have everything: a modern smartphone, stable internet, technical literacy, and access to formal recycling infrastructure. We asked a different question: what does sustainability look like for the 2.7 billion people who don't fit that profile? In Nigeria specifically, waste management sits at the intersection of three real problems. First, there is no accessible, localized tool for identifying common waste types — no existing AI model has ever been trained on sachet water nylon, okada parts, or garri packaging. Second, the majority of the population either lacks a smartphone or lives in areas with unreliable internet. Third, even people who want to recycle often don't know where to go or how to get there sustainably. EcoSnap Global was built to close all three gaps at once.

What it does

EcoSnap Global is a multi-channel waste intelligence platform that meets users exactly where they are — regardless of their device, connectivity, or technical ability.

AI Waste Scanner (Online + Offline) Users photograph or upload an image of any waste item. A machine learning model trained on Nigerian-specific waste categories — sachet water nylon, PET bottles, food waste, e-waste, polythene bags, and metal scrap — classifies the item within seconds. The result screen shows the waste type, a disposal method tailored to local infrastructure, the environmental impact in kilograms of CO₂ saved, and the nearest recycling option. Critically, the AI model runs via TensorFlow Lite directly in the browser, meaning it works with zero internet connection. For rural users and areas with poor connectivity across Nigeria and the broader continent, this is not a bonus feature — it is the product.

SMS Intelligence Mode For the significant portion of Nigeria's population using basic feature phones, EcoSnap provides a full SMS interface powered by Africa's Talking, Twilio, or Termii. A user texts a single keyword — "plastic", "nylon", "food waste" — to a shortcode and receives a tailored reply within seconds: what the waste is, how to dispose of it correctly, a localized warning (e.g. "Do NOT burn — toxic fumes"), and the nearest collection point with walking distance. The entire interaction costs nothing beyond a standard SMS rate and works on a ₦3,000 phone with no data plan.

EcoRoute Naija Once a user knows where their waste should go, EcoSnap guides them there using the lowest-carbon route available. Unlike solutions that suggest "take the subway" in a city with no subway, EcoRoute Naija is built around the transport options Nigerians actually use: walking, keke napep, okada, danfo, and private car. For each option, users see the naira cost, estimated CO₂ emissions, and travel time. The system consistently recommends walking when distance allows, reinforcing sustainable behaviour beyond just waste disposal.

Nigerian Waste Dataset The model is trained on waste categories that reflect what Nigerians actually encounter: sachet water bags (the single most littered item in Nigerian urban areas), plastic PET bottles, organic and food waste, consumer electronics and chargers, polythene shopping bags, and metal scrap including informal components like okada parts. These class names are deliberate — "pure_water_sachet" and "okada_part" are immediately recognisable to any Nigerian judge reviewing the project.

Eco Points and Leaderboard Every successful scan earns eco points. A community leaderboard ranked by city and state drives friendly competition and repeat engagement. Gamification is not decorative here — behaviour change research consistently shows that social comparison and reward loops sustain environmental habits far longer than one-time awareness campaigns.

How we built it

The architecture is designed around a single principle: one backend, multiple access channels. Every piece of intelligence — the waste classifications, disposal advice, local context, and route calculations — lives in a shared rules and routing engine that both the app and the SMS gateway query identically. Frontend is built in React with Next.js 14. The interface is designed for low-bandwidth environments: minimal assets, aggressive caching, and a progressive web app manifest so users on slow connections can install it to their home screen and use it offline. Offline AI uses TensorFlow.js with a Teachable Machine-exported model converted to TensorFlow Lite format. The model loads from the /public/model/ directory and runs inference entirely on the user's device. The image pipeline resizes inputs to 224×224 pixels, normalises pixel values, and runs classification returning both a label and a confidence score. Training used 50–100 images per class, collected locally to ensure the model reflects the actual visual appearance of Nigerian waste rather than stock photography. Backend is Node.js with Express, connected to Firebase Realtime Database for user points, scan history, and leaderboard data. Firebase Auth handles anonymous sessions for users who scan without creating an account, so there is no sign-up friction. SMS integration uses Africa's Talking as the primary gateway for its strong Nigerian network coverage and local support. The webhook endpoint receives incoming messages, normalises the text, matches keywords against the waste knowledge base, constructs a response including disposal advice and route suggestion, and returns it in the SMS-compatible XML response format. The entire SMS round trip takes under one second. EcoRoute engine is a lightweight calculation layer that takes a distance in kilometres and returns all available transport modes sorted by CO₂ emissions. Emission factors are based on published estimates: 0 kg/km for walking, 0.10 for danfo, 0.15 for keke napep, 0.18 for okada, and 0.21 for private car. These are deliberately kept as simple multipliers rather than complex routing algorithms — accuracy matters less than demonstrating the concept clearly and reliably in a demo environment. Deployment runs on Vercel. The entire application — frontend, API routes, and SMS webhook — lives in a single Next.js project, making deployment a one-command operation.

Challenges we ran into

The dataset problem. No curated, labelled dataset of Nigerian waste exists. We could not simply download a pre-trained model and apply it. Every image in our training set had to be sourced specifically for this context — photographed locally, contributed by contacts, or carefully selected from online sources with attention to visual authenticity. The classes themselves required thought: "plastic bottle" is too broad when the specific item judges will recognise is a PET water bottle with a particular label style. Getting to a working model with meaningful accuracy on locally-relevant inputs was the hardest technical challenge. Designing for multiple literacy levels. The SMS interface had to work for users who might type "nylon bag", "nylon", "nylon pure water", or just "water". A rigid keyword matcher fails them. The solution was a tiered matching system: exact phrase first, then individual keywords in priority order, then a fallback response that lists accepted keywords rather than returning an unhelpful error. Every edge case in the SMS flow represents a real user who has never interacted with a chatbot before. Offline-first architecture in a web context. Making TensorFlow Lite work reliably in a browser without any server round-trip required careful management of model loading, caching via service workers, and graceful degradation when the model hasn't loaded yet. The model file itself is several megabytes — on a slow first load, users needed clear feedback that the AI was preparing rather than experiencing a broken interface. Keeping the EcoRoute honest. It would have been easy to build EcoRoute as a simple CO₂ calculator that ignores the reality of Nigerian transport. Real routing data for keke napep and okada does not exist in any public API. The decision to use distance-based estimates with transparent emission factors, clearly labelled as approximations, was a deliberate choice: it is better to be honestly approximate than precisely wrong.

Accomplishments that we're proud of

We are most proud of the SMS mode, not because it was technically complex — it wasn't — but because it is genuinely the most important feature in the application. Building a polished, AI-powered scanning app is satisfying. Building a system that delivers the same core value to someone texting on a decade-old Nokia is meaningful. The moment we tested a live SMS and received back a formatted disposal response with a route suggestion felt different from a typical software demo. The Nigerian waste dataset is another genuine accomplishment. Every team at this hackathon will have used a pre-existing model or a generic image classification API. Our model was trained specifically on the waste that sits on Nigerian streets, in Nigerian markets, and in Nigerian homes. When a judge scans a sachet water nylon and the model correctly identifies it with over 95% confidence, that specificity is not accidental — it is the result of deliberate localisation that most teams will have skipped entirely. We are also proud of the completeness of the user journey. EcoSnap does not just tell users what their waste is. It tells them what to do, shows them the environmental consequence of doing it, guides them to where they need to go, tells them how to get there with the smallest carbon footprint, and rewards them for completing the loop. Most sustainability apps pick one of those steps. EcoSnap covers all of them.

What we learned

The most important lesson was about the distinction between localization as a feature and localization as a philosophy. Adding a Nigerian flag emoji to a generic app is a feature. Training a model on sachet water nylon, naming routes after keke napep, and designing the SMS flow around the vocabulary of a first-time feature phone user — that is a philosophy. Judges who work in impact tech can feel the difference immediately. We also learned that the constraint of the SMS interface was, paradoxically, what made the AI smarter. Because the SMS system cannot rely on an image, every waste classification had to be expressed in plain language that was immediately understandable. That discipline improved the app interface too — if a response works as a two-line SMS, it works as a push notification, a screen reader announcement, and a voice response. Accessibility thinking forced clarity. Finally, we learned that the EcoRoute concept is under-explored. CO₂ impact data for informal transport modes in developing cities simply does not exist in structured form. There is a real research and product opportunity in building ground-truth emission data for keke, okada, and tricycles across African cities. EcoSnap scratches the surface — a full solution would require partnerships with transport agencies and field data collection.

What's next for EcoSnap Global

The immediate next step is a partnership with Lagos Waste Management Authority (LAWMA) to integrate real-time collection point data and make the EcoRoute genuinely accurate rather than estimated. Beyond Lagos, the dataset expansion is the highest-leverage investment. Each new country requires a new set of locally-specific waste classes — the dominant waste types in Nairobi are different from Lagos, which are different from Accra. A federated data collection model, where local community partners photograph and label waste in exchange for access to the platform, could produce a pan-African training dataset within twelve months. The USSD channel is a priority that SMS cannot fully replace. USSD works on 2G networks with no SMS cost to the user, operates even on the most basic SIM-only devices, and is already familiar to hundreds of millions of Africans from banking and airtime top-up interactions. An EcoSnap USSD menu would unlock users that even the SMS gateway cannot reach. Voice support for low-literacy users is the long-term horizon. A user who cannot read should still be able to hold their phone near a pile of waste, hear "that is plastic — do not burn it, take it to the market on Broad Street, fifteen minutes walking" and act on it. That is the version of EcoSnap that justifies the tagline: sustainability with no barriers.

Built With

Share this project:

Updates

Submission history