Tonight’s surplus food, at a last-minute price — delivered by couriers who are free.
Restaurants, cafés, bakeries and markets box the food they will not sell today. Like a last-minute ticket, the price falls towards closing time. Couriers who are idle after the dinner rush carry it, several boxes on one route, and the app shows every customer the best value for money — food plus delivery, weighed against what the box is worth. Whatever is still unsold at closing goes to Atikova’s kitchen-waste pickup, for compost or biogas.
This is a model we are designing. The offers below are a simulation with invented sample businesses; no restaurant or courier company is a partner yet.
How it would work
- 1
The business posts its surplus
Boxes, normal value, the lowest price it accepts and when the last delivery leaves — in a minute, from the phone.
- 2
The price falls like a last-minute fare
It falls steadily from the start of the sale, never below the floor the business sets. Few boxes left — slower fall.
- 3
Idle couriers, shared routes
After the dinner rush many couriers wait for orders. The AI puts boxes going the same way on one courier, so each customer pays a share.
- 4
Search like a flight
The customer sees the best value near them, the delivery time, and “buy now or wait”: the price in 30 minutes and the chance a box is still there.
- 5
Nothing goes to the bin
What is not sold by closing is collected with the kitchen waste and goes to compost or biogas instead of landfill.
Simulator
Move the clock and pick your district. Prices, delivery and advice are calculated by the model below — the businesses are invented.
7 offers, best value first
Closed — to the compost pickup
- Fruit & vegetable rescue bag: about 7.5 kg unsold → Atikova pickup
For the business: one evening
Expected boxes sold and income for each sample offer, against throwing the food away. Estimates from the same model.
| Offer | Sold | Average price | Income | After commission | To compost |
|---|---|---|---|---|---|
| Persian rice & kebab box sample | 4 / 6 | 274 TL | 1,095 TL | 898 TL | 1.2 kg |
| Afghan Kabuli palaw box sample | 6 / 8 | 212 TL | 1,270 TL | 1,041 TL | 1.2 kg |
| Bread & pastry bag sample | 7 / 10 | 154 TL | 1,080 TL | 886 TL | 2.4 kg |
| Sandwich & cake box sample | 4 / 5 | 176 TL | 705 TL | 578 TL | 0.5 kg |
| Mixed grill box sample | 5 / 7 | 252 TL | 1,260 TL | 1,033 TL | 1.2 kg |
| Fruit & vegetable rescue bag sample | 9 / 12 | 123 TL | 1,110 TL | 910 TL | 7.5 kg |
| Home-style dishes box sample | 6 / 9 | 192 TL | 1,150 TL | 943 TL | 1.8 kg |
| Patisserie cake box sample | 4 / 6 | 238 TL | 950 TL | 779 TL | 1.4 kg |
Commission in this simulation: 18% of the food price — a figure to test in the pilot, not a decided price.
What exists in Türkiye today
Surplus-food apps already work here, all with pick-up. The gap we see: delivery by idle couriers, a price that moves with the clock, and a home for what is still unsold.
| App | Delivery | Price | What is not sold |
|---|---|---|---|
| Sepette | Pick-up at the stated hour | Surprise packages, up to 70% off | — |
| Kurtar | Pick-up from the business | Surprise box at about a third | — |
| Foodya | Pick-up with a code (no courier service) | End-of-day packages | — |
| Too Good To Go | Pick-up (not in Türkiye yet) | Surprise bag at about a third | — |
| Atikova A new model | Idle couriers, shared routes | Falls by the clock, with a floor | Compost / biogas pickup |
How good is the AI? A simulation check
We simulated 600 evenings with invented customers and offers (low, medium and high demand) and gave the same 40 decisions to our engine and to six AI language models. A decision is right when it picks the offer that was really best for that customer and gives the right “buy now / wait” advice, judged by what happened next in the simulation.
| Model | Right offer | Right advice | Quality | Harder test* | Answer time |
|---|---|---|---|---|---|
| Atikova engine v2 (on this page) | 95% | 88% | 91% | 76% | < 1 ms |
| GPT-OSS 120B20 of 40 questions (cost) | 90% | 65% | 78% | — | 15.5 s |
| Llama 3.1 8B | 88% | 57% | 72% | 60% | 3.4 s |
| Gemma 4 26B | 80% | 57% | 69% | 62% | 0.9 s |
| Qwen3 30B | 65% | 70% | 68% | 68% | 0.5 s |
| Llama 3.3 70B | 75% | 57% | 66% | 57% | 0.7 s |
| Mistral Small 3.1 24B | 75% | 50% | 62% | — | 0.9 s |
| Atikova engine v1 (first version) | 40% | 68% | 54% | 49% | < 1 ms |
| Rule: fastest delivery, always buy | 55% | 50% | 52% | 46% | — |
| Random choice | 22% | 50% | 36% | 44% | — |
- The check found real flaws in v1 and we fixed them: the buy-or-wait advice was right 61% of the time (v2: 81%); its shared-route discount rarely happened, losing about 15 TL per delivery (v2: 4.5 TL); and its price curve reached 79% of the best revenue (v2: 91%).
- At medium demand v2 sells 67% of the boxes (v1: 53%). With customers who behave unlike the engine assumes, its advice stays right 82% of the time (v1: 56%).
- Language models pick a good offer fairly often but are weak at “wait”: they mostly say “buy now”. They are also slower and cost money per question — so the engine decides, and the language model explains the choice in chat.
- Pick-up competitors sell more in the simulation (84%), because there is no delivery fee. Letting the customer choose pick-up or delivery lifts our sales to 71%; about 57% then choose pick-up.
* Harder test: simulated customers whose tastes, budgets and patience differ from the engine’s assumptions (four cheaper models only). This is a simulation built on our own assumptions, not a measurement with real customers; the real quality is only known after a pilot.
Assumptions (estimates unless a source is linked)
- Deepest discount 80% before the floor; the business’s floor always wins (the Turkish apps advertise “up to 70%”).
- One courier drop ≈ 70 TL + 12 TL per km beyond 2 km, +45 TL across the Bosphorus; a courier is paid about 38 TL per package (source).
- Up to 35% off delivery when most couriers are idle; each extra box on the route adds 25% of one drop.
- A motorcycle averages about 19 km/h in the evening; 8 minutes to pick up.
- Sales speed up as the price falls (calibrated so the samples sell about 70%, as in our simulation); the curves are illustrative, to be replaced by pilot data.
Sources
Food safety comes first
Hot dishes need temperature control and a short delivery window; bakery, cold dishes and packaged goods are simpler to start with. Hygiene rules, labelling and who is responsible for the food must be checked with the food-safety authorities before any pilot.
Would you test it?
We are looking for a few restaurants, cafés and bakeries in Istanbul to test this model with us. Tell us on the cooperation form and write “last-minute” in the message.
Restaurants: join the pilot