AgriFlow Nexus

AgriFlow Nexus is a AI powered multi-agent platform that aims to slashes SADC farm-to-market costs, predicts prices, and grades sustainability for smarter, greener logistics.

Comment

My father is a cattle farmer, his farmer was a cattle farmer too, so the plight of farmers is something that I am very passionate about. He like many farmers across southern africa lose up to 35 % of commodity value between gate and port because fuel is expensive, routes are risky, and data sits in silos. I asked: What if dozens of specialised AI “mini-experts” could co-operate live to plan routes, forecast prices, and measure CO₂ in one shot? Google’s new Agent Development Kit (ADK) gave us the perfect playground to try.

AgriFlow Nexus is a multi-agent supply-chain copilot for the 16-nation SADC bloc.

  • FieldSentinel flags drought/flood risk from 30-day soil & rain data.
  • WeatherOracle produces a 7-day harvest window forecast.
  • PricePredictor cascades BigQuery ML, rolling averages, and spot prices to forecast 12 months ahead.
  • ConflictGuard scans ACLED for violence within 50 km of planned routes.
  • LogisticsMaster solves a Vehicle Routing Problem with OR-Tools, snaps the last leg to real roads, and prices fuel.
  • SustainabilityAgent turns km + tonnes into CO₂, water use, and grades A+…F. All results surface in a Streamlit “what-if” dashboard—one click for farmers, hauliers, or regulators.

How we built it

  • Data gravity: 16 GB of public rainfall, fuel, commodity, and conflict CSVs → BigQuery (ag_flow).
  • Agents: Seven Python classes extend a 40-line BaseAgent; ADK wires them.
  • Orchestration: Thread-pooled stage 1, per-commodity PricePredictor loop, and graceful fall-backs.
  • Infra: Cloud Run container for Streamlit, Vertex AI for adaptive learning, * Secret Manager for creds.
  • Front-end: Tailwind-styled Streamlit app with Altair charts.

Challenges we ran into 1.. Shapely → OR-Tools coordinate swap bug produced 0 km routes—found in hour 18!

  1. Syncing seven agents’ schemas without breaking the orchestration tests.

Accomplishments that we're proud of a) Close to 12 % fuel reduction and 30 % faster solver time in corridor simulations. b) End-to-end ESG metrics on day 2 rarely seen in hack projects. c) A plug-and-play agent template that newcomers extended in <30 min.

What we learned

  • Modular agents beat monoliths: swapping PricePredictor for a stub kept the pipeline alive.
  • ADK’s simplicity hides real powe thread pools + fall-backs were all Python, no vendor lock-in.
  • Secrets management matters: one misplaced newline can halt your entire ML stack.

What's next for AgriFlow Nexus

  1. Real-time mode: Pub/Sub-driven alerts that trigger auto-re-planning.
  2. Reinforcement-learning NegotiationAgent for dynamic farmer–buyer pricing.
  3. Government dashboard plug-in: national-scale CO₂ & food-security metrics.
  4. Open-sourcing the agent template so any African startup can drop in its own “mini-experts”.

Built With

Share this project:

Updates