👤
What started as a hackathon project turned into something bigger. Why build another basic crop recommender when farming actually needs a full decision-support system?
When Abhiram Mukkamala, Ayaan Shaikh, and I set out to build 𝗞𝗿𝗶𝘀𝗵𝗶𝗠𝗶𝘁𝗿𝗮, we didn't want to hand in a standard demo or a thin API wrapper. Most hackathon agri-tools don't connect with the real world, because a prediction script alone doesn't solve what a smallholder farmer actually deals with day to day. So we went past the original scope and built something driven by 𝘄𝗵𝗮𝘁 𝗳𝗮𝗿𝗺𝗲𝗿𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗻𝗲𝗲𝗱:
• 𝗩𝗼𝗶𝗰𝗲 𝗔𝗜 (𝗦𝗮𝗿𝘃𝗮𝗺 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻): Smallholder farmers shouldn't have to type prompts in English. A voice-first interface in regional dialects is what actually makes this usable in the field.
• 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝘃𝗶𝘀𝗶𝗼𝗻 𝗳𝗼𝗿 𝗽𝗲𝘀𝘁 𝗱𝗶𝗮𝗴𝗻𝗼𝘀𝘁𝗶𝗰𝘀: It's hard to describe a pest pattern in words. A photo of the leaf and an instant diagnosis can save a crop before it's too late.
• 𝗔 "𝗰𝗼𝘂𝗻𝗰𝗶𝗹 𝗼𝗳 𝗺𝗼𝗱𝗲𝗹𝘀.": In agriculture, one bad AI guess can wipe out a farmer's whole season. So we built a consensus engine that cross-checks predictions and domain data across multiple models before it gives any advice.
• 𝗟𝗶𝘃𝗲 𝗺𝗮𝗻𝗱𝗶 𝗳𝗲𝗲𝗱𝘀: Growing a good yield is only half the job. Without live price data, farmers walk into the market with no bargaining power.
𝗧𝗲𝗰𝗵𝗶𝗰𝗮𝗹 𝗱𝗲𝘁𝗮𝗶𝗹𝘀:
• Predictive ML engine — Built XGBoost and Scikit-learn regressors on 7-variable soil/climate feature vectors (N-P-K, pH, temperature, humidity, rainfall) to recommend crops and fertilizer.
• Computer vision pipeline — Trained an EfficientNetB3 CNN on the IP102 dataset (.h5 artifacts) for real-time pest detection from field photos.
• Multi-model "council" orchestrator — Built a custom consensus layer that pulls outputs from local ML predictors and external LLMs (Gemini, Anthropic), cross-checking them before returning agronomic advice.
• Voice processing — Integrated Sarvam AI's APIs for audio capture, ASR, and TTS, so farmers can talk to the system in their own dialect.
• Telemetry & market feeds — Pulled live market data through the India DataGov and CEDA APIs.
• Full-stack & DevOps — Next.js frontend with geospatial mapping, risk alerts, and soil health dashboards, on an async FastAPI backend. Git LFS handles the large model files so CI/CD doesn't choke on them.
𝗦𝘁𝗮𝗰𝗸:
• ML & vision: TensorFlow/Keras (EfficientNetB3), XGBoost, Scikit-learn, Git LFS
• GenAI & audio: Sarvam AI (ASR/TTS), LangChain, ChromaDB, Gemini, Anthropic
• Backend: FastAPI (Python 3.11), Uvicorn, async routing
• Frontend: Next.js, React, Tailwind CSS, interactive mapping
𝗚𝗶𝘁𝗛𝘂𝗯:https://lnkd.in/e-jfgg-9
Take a look and let us know what you think.👇
#MachineLearning #DeepLearning #FastAPI #NextJS #ComputerVision #SystemArchitecture #AgriTech #SoftwareEngineering #Python #KrishiMitra
6 Agosto 2026 à 15h30
Ver en LinkedIn →
