KrishiAgent

Inspiration

Smallholder and marginal farmers across regions face severe agricultural vulnerabilities driven by erratic climate shifts, localized soil nutrient depletion, and fast-moving crop diseases. Traditional agricultural support systems suffer from heavy administrative bottlenecks, relying on delayed manual surveys and generalized guidelines. Inspired by the need for automated, hyper-local, and autonomous climate action, KrishiAgent was built to merge national datasets and satellite telemetry directly into actionable farm-level intelligence without requiring manual human prompting.

What It Does

KrishiAgent is an autonomous, production-grade multi-agent digital agriculture network. Operating independently in the background, it coordinates specialized sub-agents to:

• Scout Geospatial Data: Periodically ingest satellite telemetry, including Sentinel-2 NDVI vegetative indices, volumetric soil moisture, and surface temperatures via Google Earth Engine.

• Diagnose Crop Pathologies: Process plant pathology imagery using Gemini 3.5 multimodal vision capabilities to detect conditions like Yellow Rust, Bacterial Blight, and Fall Armyworm.

• Benchmark Historical Baselines: Query massive national and international datasets via Google BigQuery to contextualize yield risks against 10-year baselines.

• Deliver Vernacular Action: Translate structured advisories into multiple regional Indian languages and synthesize localized audio bulletins using Google Cloud Text-to-Speech.

How We Built It

The application relies on a robust, multi-tier Google Cloud architecture:

• Orchestration & Agent Framework: Built using the Google Antigravity SDK and Gemini 3.5 via Vertex AI, orchestrating task execution through specialized sub-agents (agent_core.py, sub_agents.py).

• Modular Tooling: Integrates dedicated Python modules for Google Earth Engine (earth_engine.py), Gemini Vision (vision.py), Google Cloud Translation & TTS (voice.py, translate.py), and Google Maps visualization (maps.py).

• Cloud Infrastructure: Persists live state to Firebase Firestore, performs historical analysis via BigQuery, and runs as a fully containerized serverless application on Google Cloud Run.

Challenges We Faced

• Asynchronous Multi-Source Synchronization: Coordinating asynchronous background loops to fetch heavy satellite matrices concurrently with real-time text translation and audio generation required careful event-loop management (asyncio).
• Robust Offline Fallbacks: Designing the system to execute cleanly even in isolated development environments by implementing intelligent synthetic data fallbacks for Google Earth Engine and Gemini credentials.

What We Learned

We mastered production-grade autonomous agent design using Google's ecosystem, moving beyond simple conversational loops into background-executed workflows. We also gained deep practical experience integrating geospatial APIs, multi-lingual translation layers, and serverless deployment pipelines on Google Cloud Run.

Built With

• firebase-firestore
• gemini-3.5-vertex-ai-google-antigravity-sdk-genai-sdk-google-earth-engine-google-maps-platform-google-cloud-speech-to-text
• google-cloud-bigquery
• google-cloud-run
• text-to-speech
• translation