Sembramos: AI-Native Wholesale Agriculture

Sembramos: Venezuela's AI-native agricultural marketplace. Gemini agents screen seller risk, price crops from real transactions, and parse voice listings — never asking prices.

Comment

Inspiration

In Táchira, farmers already have a weekly price reference — an informal number that circulates through WhatsApp groups every Sunday, set by traders nobody can quite name. But the farmers I've talked to say they almost never actually get paid that number; buyers always find a reason to pay less, and there's no way to trace where the number came from in the first place. It's not a fair price — it's just the only one anyone has, so farmers trust it because they have no alternative.

I'm not a trained developer — I learned by building this, one real problem at a time. Sembramos started before we even knew this hackathon existed, as a real answer to that exact problem in our own community, and it only exists because of the people who made the space for it: my co-founder and partner Ángeles, who tested every real release on her own phone and caught bugs I never would have found from a desk — while also carrying our family so I could build — and my daughters, who are the reason it had to work. We found Build with Gemini XPRIZE early on — and it became the fuel that pushed us to build this for real.

What it does

Sembramos is Venezuela's AI-native wholesale agricultural marketplace, live in production. The reference price for every crop is the median of confirmed, completed transactions — never asking prices — so it can't be gamed by a single listing.

AI runs the operation, not just a chat window bolted onto a form:

  • A voice/text agent (Gemini 2.5 Flash) lets a farmer who can't read or write publish a harvest by describing it out loud, resolved against a real product catalog.
  • An AI purchasing assistant builds a multi-item wholesale order, then autonomously scans every responding seller's real transaction and dispute history to produce a risk assessment before the buyer commits.
  • A planting recommendation engine cross-references real closed sales, active planting registrations, and real crop-cycle timelines — and is explicitly honest when there isn't enough data yet to recommend responsibly.
  • A crop-health tool diagnoses a photographed plant disease using Gemini + Google Search grounding, informed by anonymized patterns from other farmers who solved the same problem.
  • A full accessibility layer ("Voice Mode") lets illiterate users navigate, act, and dictate across the entire app by voice.

AI decides what to surface, when to notify, and when to flag risk — it never has authority to move money or close a contract; every financial commitment still requires an explicit human tap.

How we built it

We built Sembramos on Flutter and Dart for a single mobile-and-web codebase, backed by Firebase (Firestore, Auth, Hosting, Storage, Cloud Functions, App Check, Cloud Messaging) and Google Cloud Scheduler for autonomous jobs. All AI runs on Gemini 2.5 Flash via Vertex AI, with Google Search grounding for planting-cycle and crop-health analysis. The platform is live at sembramos.web.app with active Cloud Functions running 24/7 — not a demo environment.

Challenges we ran into

The hardest part of building AI-native wasn't calling the model — it was disciplining what it's allowed to decide. Early on, letting Gemini guess freely (units of measurement, ambiguous product variants) silently corrupted real listings; the fix was constraining its output against ground-truth catalog data and asking a clarifying question instead of guessing. Hardening Firebase App Check taught us the same lesson from the infrastructure side: only Firebase SDKs attach the App Check header automatically, so tightening security broke our own internal calls before it blocked a single bot. And building fraud protections on Firestore taught us a subtle transaction rule — a read after any write in the same transaction throws, which we found by tracing our own money-moving code.

Accomplishments that we're proud of

Sembramos is not a prototype — it's a production platform with 49 active Cloud Functions, a full bilateral contract and dispute-resolution system, reputation scoring, and organic real-user growth — built by showing up publicly on social media to teach the real problem first, with zero paid marketing. We built a complete AI purchasing assistant that autonomously screens seller risk before a buyer commits, and a full voice-accessibility layer so a farmer who can't read or write can publish a harvest and manage every order by speaking.

What we learned

Technology alone doesn't solve the problem; trust does. The reference price only works because it's built from real closed transactions, not opinions — that single design choice is what makes "see the real price before you negotiate" credible instead of just another number someone could fake. And Gemini's shared-capacity pricing (no fixed per-project quota) meant we could design "the AI does the work" as the default, not a rationed feature.

What's next for Sembramos

With the platform already live and functionally complete, our focus for the rest of the XPRIZE window is traction: converting organic users into paying transactions, expanding the wholesale buyer side (ferieros and intermediaries), and deepening predictive planting analytics as more real harvest data comes in — the recommendation engine is already built to stay honest and hold back a suggestion until the real data supports it.

Longer-term, our ambition is bigger than a marketplace: consolidating harvests through regional collection centers, connecting directly to national retail buyers, and running the entire chain — collection, quality, logistics, pricing — as a single AI-operated agricultural engine, with guaranteed fair terms for the farmer. This is a distinct, later-stage phase that requires capital we don't have yet; we're building the trust and the real transaction data first, because that's the foundation any of it stands on.

Built With

  • agentic-ai
  • agtech
  • artificial-intelligence
  • claude-code
  • cloud-firestore
  • cloud-scheduler
  • dart
  • firebase
  • firebase-app-check
  • firebase-authentication
  • firebase-cloud-functions
  • firebase-cloud-messaging
  • firebase-hosting
  • firebase-storage
  • flutter
  • gemini
  • gemini-2.5-flash
  • google-cloud
  • google-search-grounding
  • javascript
  • node.js
  • progressive-web-app
  • recaptcha-v3
  • vertex-ai
  • web-speech-api
Share this project:

Updates