Interface 01 — intake
The kitchen never posts
A closing manager will not complete a form at 23:00. Every additional field measurably reduces posting rate, and unposted surplus is indistinguishable from no surplus.
Harvest — AI that matches extra food with good people
AI4Good moves surplus food from grocers and restaurants to the people who need it, before it spoils.
Eighteen million American households go short on food. Seventy million tons of perfectly good food are lost along the supply chain. The food and the hunger sit in the same neighborhoods on the same night. Matching them inside the shelf life means weighing too many variables, against too short a clock, too many times a day for anyone to do by hand — which is exactly what AI is for.
of food goes unsold or uneaten in the US annually — roughly 29% of the supply.
ReFED 2026of US households were food insecure in 2024; 18.3 million homes.
USDA ERS ERR-358car-years of greenhouse gas emissions attributable to wasted food.
ReFED 2026tons a year recoverable through donation matching alone, at $595M net benefit.
ReFED solution analysisWatch — 90 seconds
From the gap, to the AI that closes it, to what it means for a grocer’s bottom line — in a minute and a half.
Playing muted — tap the speaker icon for sound.
Why the food does not arrive
Every link in the food chain leaks. Farms leave crops unharvested, manufacturers over-run, distributors hold short-dated pallets, grocers take shrink, kitchens over-prep. There is enough food and there are people who need it. What breaks is the handoff — three times over, and each break ends the same way: a family eats less, and good food goes in a dumpster.
Interface 01 — intake
A closing manager will not complete a form at 23:00. Every additional field measurably reduces posting rate, and unposted surplus is indistinguishable from no surplus.
Interface 02 — allocation
Bulletin-board tools award food to whoever refreshes fastest, not to the site that can store, use and distribute it within its remaining window.
Interface 03 — last mile
Phone trees, group texts and a volunteer deciding drop order in a parking lot. Perishables lose their window in transit, and nobody logs the temperature.
How a meal actually happens
Harvest Donor
A photograph is the input. Vision classifies the item, estimates mass, flags the top-nine allergens and initializes the safe-handling clock.
Harvest Dispatch
Multi-stop runs with time windows, refrigeration constraints and drop order resolved before a driver accepts. Photo handoff at both ends.
Harvest Table
Capacity, refrigeration, service hours, population served and dietary requirements, declared once and matched against continuously.
The AI, and what it actually does
Every hour a pan of food sits unclaimed is an hour closer to being worthless to anyone. The engine prices that clock and keeps re-solving who should get the food, rather than posting it and hoping somebody claims it in time.
Vision classification plus entity normalization maps free-text kitchen shorthand onto a canonical taxonomy aligned to USDA FoodData Central, so downstream models can reason about protein, allergens and storage class.
Remaining safe window inferred from category, storage state, holding temperature and elapsed time since preparation. This value is the constraint every later stage is scored against.
Per-site 72-hour projection by food category from historical draw-down, day of week, SNAP issuance timing, weather and local events. Demand is non-stationary; week three does not resemble week one.
Objective maximizes nutritional value delivered, dietary and cultural fit and equity weighting toward underserved tracts; penalised by travel time, spoilage probability and storage overflow. Solved as a rolling minimum-cost flow.
Pickups and drops bundled into multi-stop runs under time-window and refrigeration constraints. Route density is the primary lever on cost per pound.
Every assignment emits a plain-language rationale covering food, site, driver and window. Food-safety decisions are not permitted to be opaque, and an unauditable recommendation is not trusted twice.
Declines, no-shows and post-delivery utilization reports are fed back into the demand forecaster, which improves precisely where precision matters most: predicting what a specific community will actually consume. Open the match console →
Why restaurants and grocers say yes
Goodwill gets a program started. Arithmetic keeps it running after the manager who championed it moves on. Modeled here for one mid-size grocery location moving about 1,200 lb of recoverable surplus a week — roughly 52,000 meals a year, and cash-positive for the store.
Who makes a meal happen
Donor
Agency
Driver
How the food moves
Cost per pound is mostly route density and driver cost. One driver pool cannot cover both the cheap hours and the expensive ones, so we do not use one. Capacity is assembled in tiers, and the engine sends each run to the least costly tier that can still reach the table in time.
Delivery platforms already run food-recovery programs. DoorDash’s Project DASH reports more than 150 million meals distributed across over 9 million deliveries since 2018, with the partner organization deciding who receives food, when, and from where. Where a leg is scheduled and predictable, that capacity exists already and the driver never leaves their own platform — which means their insurance and their status travel with them.
Surplus appears at closing time. Agencies open in the morning. Fire stations are staffed around the clock, sited by response-time geography so they already sit where people are, and understood as somewhere you can walk up to without being asked anything. A late drop waits there safely until morning. Never the apparatus — an engine is dispatchable at any second, so it is the building we need, not the truck.
Off-duty firefighters and rideshare drivers, in the hours that are already dead for earning — midweek afternoons between the lunch and dinner rush, weekday mornings once the commute clears. The engine offers runs that end where demand is, so a driver is repositioned on a trip they were going to make empty. We do not ask anyone to give up a shift.
For what nothing else covers: restaurant close, cold chain, windows too short to wait for a volunteer to accept. Fully insured, properly contracted, and deliberately small. How small it stays is the honest measure of whether the other three are working.
Every driver who moves food for us may take a box for their own household, on any day they run — no questions, no qualification, and never in place of pay. A great many drivers are food insecure themselves. Nobody who moves food for AI4Good goes home without food.
Where this actually stands. AI4Good is pre-launch, and none of these delivery relationships are in place yet. Project DASH is an existing DoorDash program we intend to apply to, not a partnership we hold, and we are not affiliated with DoorDash, Uber or Lyft. This is the operating design rather than a description of current operations — published in this much detail so that a partner can hold us to it.
Timing
Statutory
The Food Donation Improvement Act, signed 5 January 2023, extended Bill Emerson Good Samaritan Act protection to food businesses donating directly to individuals, and to sales at no more than the cost of handling and transport. The nonprofit intermediary is no longer a legal requirement.
Regulatory
State organic-waste statutes — California SB 1383 among them — increasingly require large generators to divert edible food to recovery rather than landfill, with recordkeeping obligations attached. Compliance requires systems.
Technical
Reliable food recognition from a handset, per-site demand forecasting and constrained routing were research problems a decade ago and a rounding error on infrastructure spend today. The prior constraint was unit cost, not capability.
Where we feed people next
A city covered properly feeds more people than twenty covered thinly, because route density is what drives the cost of a meal down. Charleston first, and all the way.
We start at the retail and foodservice end of the chain, because that is where surplus is most perishable and most local. The same engine runs upstream — distribution centers, manufacturers, farm-level gleaning — as coverage deepens.
120 donor locations, 45 agencies, one anchor grocery chain. 1.4M lb recovered, approximately 1.17M meal equivalents.
Four metros, 850 donor locations, 220 agencies. Launch playbook documented and transferred to a market operations team. 11.5M lb.
3,400 donor locations, 780 agencies, 48M lb. Direct-to-individual matching enabled where agency coverage is insufficient.
No proof-of-need gate. No recipient location tracking. No photographs of people. Records purged on a fixed schedule and never shared with funders.
Get involved
And there are families a few minutes away who could eat it. Founding partner slots are open in the Charleston metro, with a waitlist for Atlanta and Nashville. Fifteen minutes and one hauling invoice is all it takes to start.