AgroMint
Small farms in Azerbaijan (2–20 hectares) are primarily managed by family farmers who lack formal agronomic training. They base decisions regarding fertilization, irrigation, and pest control on three unreliable sources: random online advice, the trial-and-error experiences of neighbors, and agronomist consultations—costing around $15–25—that require scheduling weeks in advance. Consequently, fertilizer dosages are selected without considering soil analysis or the crop's growth stage. Estimates suggest this can lead to a 20–30% loss in seasonal yield. Farmers often attribute these losses to "bad weather," failing to recognize that the root cause lies in their decision-making process. Existing general-purpose AI chatbots do not improve the situation; instead, they exacerbate it. They provide identical advice for soils in Ganja and Lankaran, fail to cite sources, and ultimately cannot earn the farmer's trust. AgroMint bridges this gap. Farmers input their district, field size, crop type, and growth stage (or upload laboratory soil analysis results) to receive tailored recommendations backed by scientific sources, rather than generic, templated text.
Preliminary score
- Value for the user17.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI19.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing13.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility11.0 / 15
Data requirements, running costs and a clear next step.
- Originality6.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary65/100
- Strengths
- Your intake-to-dossier flow combines field data, live Open-Meteo 7-day forecasts and retrieval over transcribed Azerbaijani agronomy textbooks, which produces locally grounded advice rather than a generic chatbot answer. Your testing write-up is unusually honest and specific. It documents three real failures: cotton defoliation triggered by substring matching, the 'azot'/'ot' language collision, and hallucinated kg/ha doses. Each comes with a concrete fix. The code-level sanitizer that strips quantitative rates when no lab data exists is a thoughtful safety design for chemical advice. Your per-dossier cost estimate and the existing AgroSphere sales channel make the path to market credible.
- To improve
- Your user is described three different ways: small 2–20 ha family farmers in the submission, 40 ha 'Uncle Ali' in the deck, and 2,000–50,000 ha holdings in the README. Pick one and size the value and pricing for that user. Your ROI figures (20% input savings, 5% yield protected) are modeled assumptions. Even one comparison against a real agronomist's plan for the same field would make them believable. The 79 tests and 16-case benchmark are mentioned, and the Gemini-vs-rules-vs-RAG comparison is described in your docs, but the materials show no results table. Also show how much Gemini adds over the rules engine, since the app claims full functionality without an API key, which makes the AI look optional.
- GPT-6 AstraPreliminary69/100
- Strengths
- Your intake and dashboard screenshots make the field-to-advice workflow concrete, while localized agronomic sources and weather data provide useful domain context. Your testing write-up identifies specific failures—including premature cotton defoliation and the azot/ot collision—and explains targeted fixes. Your documented dosage sanitizer, simulated-weather labeling, and agronomist sign-off requirement are sensible safeguards for consequential recommendations.
- To improve
- Your materials alternate between family farmers and enterprise agronomists; select one primary workflow and measure time to an approved plan against the current process. Demonstrate a live Gemini run showing retrieved passages, citations, weather inputs, and explicit separation from the rules-only fallback. Publish per-case results for the 16 scenarios across plain Gemini, rules-only, and RAG, including expert judgments and remaining failures, since API-free test passes do not establish live-model safety. Validate actual operating costs and corpus permissions, then define a small supervised pilot to test the modeled benefits.
The team
Xanlar İsmayılov
designer
Həsən Məmmədov
data
Sevinc Muxtarova
business
Bayram SüleymanovLead
product
Sadiq Mustafayev
developer