One photo.
A harvest,
sold fresh.
Benguet grows 80% of Metro Manila's vegetables — and up to half rots before it sells. Ani grades a harvest from a single photo, scores its spoilage urgency, and matches it to live NCR demand. Three AI agents, running privately on one AMD Instinct MI300X.
It isn't a growing problem. It's a coordination problem.
Farmers harvest blind into gluts, prices crash, and produce is dumped — while there's no cold chain to hold what does sell. The waste is happening this season. Software can fix it now.
A cold-chain gap at national scale
Benguet feeds the country's largest market, Metro Manila — and the same grading-plus-matching engine extends anywhere harvest timing and demand are disconnected.
Farmer cooperatives
Better prices and fewer losses on every harvest cycle.
Traders & wholesalers
A demand-matched feed instead of guesswork ordering.
Agri-logistics operators
Coordinated dispatch across farms instead of one truck at a time.
Benguet → Metro Manila today. Southeast Asian cold chains next.
Gross-benefit estimate on cited loss rates (DA ~30% general; 42–50% vegetables) — adjust the inputs. It excludes Ani service, hardware, and operating costs, so it is not net ROI or a measured pilot result; see BUSINESS.md.
Watch three agents turn a photo into a sale
This is the real product loop. Pick a harvest and run it — the grader, the demand oracle, and the router each do their job in seconds, all on one MI300X.
Three tiers. Two env vars. One card.
The UI you just used is only Tier 1. Behind it, a thin contract lets us run against a stub today and the MI300X at showtime — swapping backends is a single variable, never a rewrite.
The product
Next.js on a Hugging Face Docker Space. Captures the harvest, streams the agent trace, renders the result.
The agentic brain
FastAPI orchestrates the 3 agents behind one contract. Backend = stub · fireworks · mi300x.
One MI300X
vLLM on ROCm serves the Gemma 3 vision grader while FastAPI keeps grading, matching, and dispatch on the same private MI300X host.
Every layer degrades gracefully to a stub — the demo never hard-fails in front of a judge.
Measured on one AMD card
The committed ROCm receipts verify the hardware and model load. Ani reports measured values only — no simulated telemetry and no unverified accelerator comparisons.
The deployment keeps the Gemma vision grader, orchestration, pricing data, and farmer photos on one private AMD host. Measured ROCm logs — not animated estimates — support the Best AMD-Hosted Gemma entry.