Inspiration
A farmer decides what to sow once a year. Once the seed is in the ground, the calendar is fixed — if a crop's critical growth stage lands during peak water stress and withdrawal restrictions kick in, the loss is already locked in. Nobody making that call today has an easy way to see, before sowing, whether this year's forecast puts a specific crop's most vulnerable week on a collision course with the water shortage.
We also noticed the hackathon's own datasets (nyc-taxi, showcase-ecommerce)
ship with planted freshness and quality problems — a deliberate playground for
exactly the kind of agent we wanted to build: one that treats a data
pipeline's health as something to actively monitor and act on, not just read.
What it does
Terroir Context Agents compares three crops on a real parcel before sowing, using live public data — no synthetic fixtures required for the primary path:
- The parcel — a farmer enters a commune; the app fetches real RPG parcel boundaries (IGN), soil texture (ISRIC SoilGrids, with a documented fallback chain), and nearby water-monitoring stations (Hub'Eau), all through public, keyless APIs.
- The result, in one screen — decision first, everything else one click away: the three crops ranked by water-stress risk and margin (€/ha), a plain-language verdict, the exact dates where a crop's critical growth stage overlaps the forecast water-tension window, and concrete corrective levers (shift sowing date, earlier-maturing variety, secure irrigation rights) each with a quantified marginal gain in €/ha. The crop-vs-crop comparator, calendar, price simulator and report sit behind two compact buttons instead of crowding the decision.
- A trust gate that refuses to lie. Every number traces to a DataHub URN. When a critical source is stale or missing, the app doesn't quietly degrade — it says so, in plain language, or refuses to produce a number at all. The full data-quality certificate, confidence dashboard, and DataHub lineage graph stay one click away, never hidden, just not forced on a farmer who wants an answer in ten seconds.
- A Sentinel agent that watches its own foundations. Simulate a station
outage and watch the invalidation propagate through the real DataHub
lineage graph —
hubeau_hydrometrie → features_bilan_hydrique → scenarios_cultures → recommandations_parcelle— tagging every downstream recommendation as unsafe and filing an incident, automatically. - A second, domain-agnostic mode for judges. The same Sentinelle agent
also runs against the hackathon's own
nyc-taxidatapack (make demo-generic), detecting the planted freshness anomaly and tagging it — proving the architecture isn't agriculture-specific, just instantiated there.
Log in or sign up for Devpost to join the conversation.