Ajwai grades incoming produce with computer vision, predicts each lot's remaining shelf life, and forecasts depot-level demand — so distributors sell the right stock, to the right buyer, before it turns.
MVP in development · Beta opening to selected distributors
A food distributor buys hundreds of crates a week with no reliable read on what condition they arrived in. Quality is judged by eye, in seconds, by whoever is on the loading bay. Stock is then dispatched roughly in the order it was stacked.
By the time a crate is visibly bad, the money is already gone — and the buyer who would have taken it two days earlier bought from somebody else.
Mid-size distributors, market wholesalers, and cold-store operators moving fruit, vegetables, and other short-life goods between farms and urban buyers.
Every discarded crate is paid-for stock, paid-for transport, and paid-for storage written off at once — on categories where margins are thin to begin with.
Unsold stock forces panic discounting, which distorts pricing for the next cycle and damages the buyer relationships that make consistent volume possible.
Ajwai turns intake into data. A phone or fixed camera photographs each lot; the model grades it, estimates remaining sell window from that grade plus storage and transit conditions, and ranks the whole warehouse by urgency. Demand forecasting then matches the most urgent stock to the buyers most likely to take it now.
Consistent quality scoring that does not depend on who is on shift.
Hours of sell time remaining, updated as conditions change.
A live dispatch order across every lot in storage.
Demand forecasts that point urgent stock at the right customer.
Ajwai is not a spreadsheet with an assistant bolted on. The product does not exist without the models — grading, shelf-life estimation, and demand forecasting are the platform.
Intake photographs of each lot, category and variety labels, supplier and origin records, transit duration, storage temperature and humidity readings, historical sale timing, and per-buyer order history.
Surface defects, bruising, mould, colour and ripeness stage, size distribution, and lot-level grade consistency from images captured in ordinary warehouse lighting.
Remaining sell window in hours per lot, spoilage risk under current storage conditions, and expected demand per product, per depot, per day.
A ranked dispatch order, buyer-matching suggestions, markdown timing recommendations, and a written daily intake report summarising quality and risk across the warehouse.
Convolutional and vision-transformer image classifiers for grading and defect detection; gradient-boosted and sequence models for time-series demand forecasting; survival-style modelling for shelf-life estimation; a large language model for report generation and natural-language querying of warehouse state.
Grading has to return a result while the truck is still being unloaded. GPU inference keeps per-image latency low enough for the bay, and GPU training lets vision models be retrained per crop, per season, and per region as new labelled intake data arrives.
Photograph a lot on any phone and get a grade, defect breakdown, and confidence score before it is stacked.
Every lot carries a live countdown that adjusts to storage temperature and time already spent in transit.
A single ordered list of what leaves the warehouse next, refreshed as new stock arrives.
Expected volume per product and depot, so buying decisions stop being made on last week's feeling.
Urgent stock surfaced against the buyers whose order patterns suit it, with suggested pricing.
Written and exportable records of what spoiled, where it came from, and which suppliers underperform.
Warehouse staff photograph each incoming lot in the Ajwai app and scan or enter the supplier reference. No new hardware is required to start.
The vision model returns a grade, defect breakdown, and an initial shelf-life estimate within seconds, attached to the lot record.
The dashboard reorders the floor by urgency. Supervisors are alerted when a lot crosses into its final sell window.
Sale and loss outcomes feed back into the models, so grading and demand forecasts improve for that crop, region, and season.
The operations view is the product's centre of gravity: one screen a supervisor can read from across the room, showing what came in, what condition it is in, and what has to go out today.
Interface preview from the MVP in development. Figures shown are illustrative sample data.
Bars show percentage of sell window remaining, averaged per category.
Ajwai runs on cloud infrastructure for image storage, model inference, API hosting, real-time alerting, analytics, and user management. Our planned AWS architecture:
| Service | What it does for Ajwai |
|---|---|
| Amazon S3 | Stores intake images and labelled training sets, with lifecycle rules moving older captures to cheaper tiers. |
| Amazon RDS | Holds lot records, supplier and buyer data, sale outcomes, and the audit trail behind every prediction. |
| AWS Lambda / ECS | Runs the grading API and scheduled forecasting jobs, scaling with intake volume rather than sitting idle overnight. |
| Amazon SageMaker | Trains, versions, and serves the grading and shelf-life models, and hosts retraining as regional data accumulates. |
| Amazon Bedrock | Powers the written daily intake reports and natural-language querying of warehouse state. |
| Amazon CloudFront | Serves the dashboard and app assets quickly to depots on constrained connections. |
| Amazon Cognito | Manages depot-level accounts and role separation between operators, supervisors, and owners. |
| Amazon CloudWatch | Monitors inference latency, job failures, and model drift so grading quality is observable in production. |
Grading is a computer vision workload, so accelerated computing is on the critical path rather than a nice-to-have.
| Technology | Where it applies |
|---|---|
| CUDA | Training the grading and defect-detection models on labelled intake imagery. |
| TensorRT | Optimising those models for low-latency inference so results return while unloading is still in progress. |
| Triton Inference Server | Serving multiple crop-specific models behind one endpoint as the catalogue grows. |
| Jetson | Planned edge deployment for fixed bay cameras at depots with unreliable connectivity. |
Food distribution across Africa is consolidating: buyers want consistent quality and reliable delivery, and distributors who cannot supply it lose contracts to those who can. At the same time, smartphone cameras in every warehouse and affordable GPU inference in the cloud have made per-lot quality assessment cheap enough to run on ordinary stock, not just premium export shipments.
The category has been managed by intuition because measuring it was uneconomic. That constraint has lifted, and the operators who instrument their intake first will hold the margin advantage.
Mid-size fresh-produce distributors, market wholesalers, cold-store operators, agro-processing buyers, and grocery and restaurant supply chains.
Nigeria first, with expansion planned across Ghana, Kenya, Rwanda, and South Africa as depot density supports it.
Per-depot subscription, tiered by monthly intake volume, with usage-based AI grading credits and enterprise plans for multi-depot groups.
Land single depots, expand across a distributor's network, then extend the buyer side into a demand marketplace built on accumulated quality data.
A working grading demo is live on this site. It runs a full measurement pipeline in the browser — background separation, colour maturity, defect detection, and temperature-adjusted shelf-life estimation — and returns a grade and sell window for any produce photo you give it.
Next is the trained model. We are recruiting a small number of Lagos depots as design partners to contribute labelled intake imagery and test the grading flow against their own stock. Demand forecasting is specified and scheduled behind that work.
Try the live grading demo Join the beta list| Phase | Milestone |
|---|---|
| Phase 1 | MVP: intake capture, grading model, and shelf-life scoring for core vegetable categories. |
| Phase 2 | Design-partner depots: labelled intake data collection and per-crop model training. |
| Phase 3 | Demand forecasting and buyer matching released to beta accounts. |
| Phase 4 | Full AWS production deployment with monitoring, retraining pipelines, and role-based access. |
| Phase 5 | GPU-accelerated inference at scale and edge grading for low-connectivity depots. |
| Phase 6 | Multi-depot enterprise tier and expansion into additional African markets. |
A small technical team based in Lagos, building in the open. The repository is public and the grading demo runs on this site.
Software developer and computer science student at Kwara State University, working across AI evaluation and applied machine learning. Has built and shipped several production web products, and now leads Ajwai's architecture, the grading pipeline, and the shelf-life model. Started Ajwai after looking at how much perishable stock is written off in Nigerian distribution for want of a measurement anyone can act on. LinkedIn
Leads model development for Ajwai — the computer vision grading model, defect detection, and shelf-life estimation, including training against labelled intake imagery collected with design-partner depots.
Leads depot engagement and product operations — recruiting design partners, running the intake workflow with them, and translating what happens on the loading bay into product requirements.
Tell us what you move and where you store it, and we will walk you through a grading demo on your produce. Beta places are limited while the models are being tuned per crop.
Or email us directly at hello@ajwai.site
We collect only what a prediction or an account needs, and customer operational data belongs to the customer. Read the Privacy Policy.
Grades, shelf-life estimates, and forecasts are decision support, not certifications. Read the Terms of Use.
Encrypted storage, role-based access, dedicated identity management, and logged infrastructure access. Ajwai does not replace statutory food safety inspection.