"Predict with Satellites. Protect with Science. Prosper with Markets."
A neuro-symbolic Generative AI agricultural intelligence platform that thinks like a senior agronomist and speaks like a village elder — fusing live satellite telemetry, official APMC economics, and scientific ICAR RAG into grounded, evidence-backed generative farm advisories.
Live Application: https://agro-shield-rho.vercel.app/
Demo Video: https://youtu.be/u1vQJKROZCY
Track: Climate, Resource Infrastructure & Global Supply Resilience
Indian agriculture sustains 140 million farmers, but smallholders face three compounding failures under extreme climate volatility:
┌─────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ THE 3 SYSTEMIC AGRICULTURAL VULNERABILITIES │
├───────────────────────────────────┬───────────────────────────────────┬─────────────────────────────────┤
│ 1. Hyperlocal Climate Shocks │ 2. Mandi Price Exploitation │ 3. The LLM Hallucination Trap │
│ │ │ │
│ • ₹40,000 Cr annual crop loss │ • 20%–35% lost in panic distress │ • Generic LLMs invent fatal │
│ from sudden unseasonal rain │ selling at harvest peak │ 10x toxic pesticide dosages │
│ • Broadcast weather forecasts │ • Farmers sell blindly due to zero│ • AI chatbots confuse crop │
│ lack field-level GPS context │ forward 14-day price outlooks │ stages (urea during boll shed)│
│ • Farmers don't know whether to │ • Government portals show only │ • English jargon is unreadable │
│ hold fertilizer or drain canals │ yesterday's historical rates │ for low-literacy smallholders │
└───────────────────────────────────┴───────────────────────────────────┴─────────────────────────────────┘
Why Generic AI Fails in Farming: A standard chatbot asked "My cotton leaves have red spots, what spray should I use?" will hallucinate chemical cocktails without knowing the crop's sowing date, current rainfall, or thermal stress. In agriculture, an AI hallucination destroys an entire season's livelihood.
AgroShield delivers 7 integrated, production-grade capabilities designed specifically for rural agricultural decision-making:
┌─────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ AGROSHIELD CORE CAPABILITIES │
├───────────────────────────────────┬───────────────────────────────────┬─────────────────────────────────┤
│ Farmarian Agronomic AI Advisor │ 14-Day Mandi Price Radar │ Multimodal Crop Clinic │
│ • Bilingual vernacular reasoning │ • Real-time AGMARKNET APMC data │ • Gemini Vision leaf diagnosis │
│ • Timeline-bounded action plans │ • 14-day forward price curve │ • Pathogen & pest identification│
│ • Native Web Speech API audio TTS │ • MSP benchmark reference floor │ • Non-chemical botanical sprays │
│ • Millisecond latency audit trace │ • Net freight realization profit │ • Legal safety & KVK disclaimers│
├───────────────────────────────────┼───────────────────────────────────┼─────────────────────────────────┤
│ Climate "What-If" Simulator │ Hyperlocal Weather Telemetry │ Multi-Crop Phenology Tracker │
│ • Interactive temp & rain sliders │ • GPS-exact Open-Meteo satellite │ • Days After Sowing (DAS) clock │
│ • Recalculates biological stress │ • 7-day synoptic rain probability │ • Phenological stage detection │
│ • Counterfactual resilience plans │ • Evapotranspiration & heat index │ • Seamless multi-crop switching │
└───────────────────────────────────┴───────────────────────────────────┴─────────────────────────────────┘
- Bilingual Colloquial Synthesis: Formulates personalized, question-specific recommendations in natural Devanagari Hindi and Indian English.
- Timeline-Bounded Action Plans: Every advisory provides structured guidance across 3 explicit timeframes:
[Immediate Step - Next 24 Hours](e.g., withhold nitrogen application before heavy rain).[Preventative Measure - Next 48-72 Hours](e.g., install yellow sticky traps at 10-12/acre).[Field Monitoring - Daily Inspection](e.g., scout 20 random plants for bollworm flare-ups).
- Audio Read-Aloud (TTS): Native Web Speech API audio synthesis in Hindi (
hi-IN) and Indian English (en-IN) for farmers with limited literacy. - Transparent Execution Trace: Full millisecond latency audit log displaying each internal agent step and tool invocation duration.
- Live AGMARKNET Market Feeds: Real-time commodity arrival volumes and APMC price spreads across 10+ crops (Cotton, Soybean, Wheat, Rice, Chickpea, Maize, Mustard, Potato, Onion, Sugarcane).
- 14-Day Forward Projection: Predictive price trend curves plotted against the official government Minimum Support Price (MSP) benchmark floor.
- Net Realization Freight Calculator: Deducts round-trip road distance (km) and vehicle freight rates (₹/km) to calculate true net farmgate profit across nearby APMC mandis.
- Actionable "SELL" vs "HOLD" Intelligence: Eliminates harvest-season distress selling by informing farmers whether holding for 7-10 days yields higher net profit.
- Instant Photo Inspection: Farmers snap and upload field photos of diseased leaves, infested bolls, or wilting stems.
- Gemini Vision Pathogen Parsing: Visual AI inspects chlorosis patterns, fungal spore lesions, and sucking pest signs.
- Biological Remediation First: Recommends non-chemical ICAR remedies (5% NSKE neem oil spray, pheromone lures) before chemical pesticides, accompanied by Krishi Vigyan Kendra (KVK) confirmation disclaimers.
-
Interactive Stress Perturbation Sandbox: Farmers and extension officers slide controls to perturb microclimate conditions:
- Temperature anomalies:
$-5^\circ\text{C}$ to$+5^\circ\text{C}$ - Precipitation shocks:
$-40\text{mm}$ to$+50\text{mm}$
- Temperature anomalies:
- Dynamic Risk Recalculation: The deterministic risk engine instantly re-computes heat stress, flower bud shedding rates, and fungal favorability, prompting Gemini to formulate counterfactual mitigation plans.
-
GPS-Exact Telemetry: Pulls real-time meteorological data directly from Open-Meteo satellite feeds at exact farm coordinates (temperature, relative humidity, wind speed, solar radiation, evapotranspiration
$ET_0$ ). - 7-Day Synoptic Outlook: Tracks daily precipitation probability and rainfall volume to prevent fertilizer runoff and guide irrigation schedules.
- Multi-Crop Support: Seamlessly switch between cultivated crops (e.g., Bt Cotton, Soybean, Desi Cotton, Wheat) on a single farm profile.
- Phenological Stage Tracking: Automatically computes Days After Sowing (DAS) to identify precise biological stages (e.g., Day 77: Flowering & Square Formation), ensuring advisories correspond to biological realities.
- Clean Accessibility: High-contrast rural daylight aesthetic (Cornsilk
#FEFADCand Coffee#241909) designed for sunlight visibility on field smartphones. - 1-Click Interactive Evaluation: Single-click demo bypasses login forms for rapid evaluator testing.
| Feature Module | Problem Addressed | Real-Time Data Source | Core Algorithmic Technique | Output to Farmer |
|---|---|---|---|---|
| Farmarian Advisory | Confusing text & generic tips | Open-Meteo + ICAR Knowledge Base | Constrained Decoding ( |
Structured 24h/48h action plan + Audio TTS |
| Mandi Price Radar | Distress panic selling at APMC | data.gov.in AGMARKNET Daily API | 14-day price forecasting + MSP regression | Interactive price chart + "SELL vs HOLD" signal |
| Freight Calculator | Hidden transport cost cuts profit | OpenStreetMap / Haversine distance | Dynamic freight deduction per quintal | True net farmgate realization (₹/Quintal) |
| Disease Clinic | Misidentifying crop infections | Multimodal image upload | Gemini Vision leaf lesion parsing | Pathogen etiology, NSKE remedy & disclaimer |
| What-If Simulator | Unpreparedness for sudden shocks | User-controlled climate perturbation | Deterministic GDD & moisture stress calculus | Forward-looking counterfactual risk advisory |
| Phenology Engine | Wrong chemical timing in season | Sowing date + Weather history | Thermal Growing Degree Days (GDD) math | Live biological stage (e.g., Day 77 Flowering) |
| Weather Telemetry | Inaccurate district-wide news | Open-Meteo Satellite Models | Hourly & daily synoptic precipitation math | 7-day GPS-exact rain, heat index & humidity |
AgroShield significantly mitigates hallucinations through a hybrid Neuro-Symbolic architecture: deterministic Python biological modeling handles the arithmetic, while Google Gemini handles multimodal perception, semantic synthesis, and vernacular reasoning.
flowchart TD
classDef input fill:#FEFADC,stroke:#241909,stroke-width:2px,color:#241909;
classDef symbolic fill:#F6D4BA,stroke:#E07B39,stroke-width:2px,color:#241909;
classDef rag fill:#E8F5E9,stroke:#3A7D44,stroke-width:2px,color:#241909;
classDef neural fill:#E3F2FD,stroke:#1976D2,stroke-width:2px,color:#241909;
classDef output fill:#241909,stroke:#FEFADC,stroke-width:2px,color:#FEFADC;
A["Farmer Query / Leaf Photo / Voice"]:::input --> B["FastAPI Agent Orchestrator"]
subgraph SYMBOLIC ["1. Symbolic Tier (Deterministic Mathematics)"]
C1["Open-Meteo Satellite Telemetry<br/>Temp, Humidity, Rain, ET0"]:::symbolic
C2["data.gov.in AGMARKNET<br/>Daily APMC Modal & Arrivals"]:::symbolic
C3["Phenological Stage Calculus<br/>Days After Sowing (DAS) & GDD"]:::symbolic
C1 & C2 & C3 --> D["Deterministic Risk Engine<br/>Computes 0-100 Stress Score & Factor Triggers"]:::symbolic
end
B --> SYMBOLIC
subgraph RAG ["2. Knowledge Retrieval Tier"]
E["Curated ICAR Reference Bulletins"]:::rag --> F["384-Dim Hashed Vectorizer<br/>Sub-1ms CPU Vectorization"]:::rag
F --> G["Neon PostgreSQL + pgvector<br/>Sub-5ms HNSW Cosine Index Search"]:::rag
end
D --> RAG
subgraph NEURAL ["3. Neural Tier (Generative AI Reasoning)"]
H["Gemini 3.5 Flash Lite"]:::neural
H1["Multimodal Vision API<br/>Leaf Lesions & Pathogen Parsing"]:::neural
H2["Constrained JSON Decoding<br/>T=0.2 Strict Schema Enforcement"]:::neural
H3["Colloquial Vernacular NLG<br/>Everyday Devanagari Hindi & English"]:::neural
H --> H1 & H2 & H3
end
D & G --> H
H --> I["Grounded Vernacular Action Plan<br/>• 24h Immediate Action & 48h Prevention<br/>• Verified ICAR Paper Citations<br/>• Millisecond Latency Audit Trace<br/>• Web Speech API Audio Synthesis"]:::output
- Separation of Concerns: The LLM is never asked to do math, calculate growing degree days, or estimate moisture stress.
-
Pre-LLM Fact Ingestion: The deterministic risk engine evaluates mathematical thresholds first:
$$\text{GDD} = \sum \max\left(\frac{T_{\max} + T_{\min}}{2} - T_{\text{base}}, 0\right)$$ -
Grounded State Vector: Only verified numerical vectors (
Risk: HIGH (74/100),Stage: Flowering,Precipitation: 38mm sum) are passed into Gemini's prompt context, guaranteeing mathematical consistency.
- Knowledge Corpus: Curated reference dataset of foundational agronomic guidance from ICAR-CICR (Central Institute for Cotton Research, Nagpur) and ICAR-IIWBR (Indian Institute of Wheat and Barley Research, Karnal) covering moisture stress, terminal heat, and pest management.
-
Embedding Strategy: 384-dimensional multi-pass n-gram hashing vectorizer running in
$<1\text{ms}$ on CPU with zero heavy external PyTorch dependencies, allowing fast, deterministic cloud deployment without binary bloat. -
HNSW Vector Retrieval: Queries the Neon Cloud PostgreSQL vector store using cosine distance, filtering by crop and prioritizing matching growth stages to return the top 3 relevant chunks:
SELECT c.clean_text, c.page_number, d.title, (1 - (e.embedding <=> CAST(:query_vector AS vector))) AS similarity_score FROM document_embeddings e JOIN agricultural_document_chunks c ON c.id = e.chunk_id JOIN agricultural_documents d ON d.id = c.document_id WHERE (c.crop_name ILIKE :crop_name OR c.crop_name IS NULL) ORDER BY (CASE WHEN c.growth_stage ILIKE :stage THEN 0 ELSE 1 END) ASC, e.embedding <=> CAST(:query_vector AS vector) ASC LIMIT 3;
- Citation Provenance: Every synthesized tip links back to the source research document, issuing institution, and reference page number.
- Visual Foliage Inspection: Ingests base64-encoded leaf/stem photos directly into Gemini.
- Structured Pathogen Parsing: Identifies necrotic spots, chlorosis patterns, sucking pest signs (curling margins), or fungal lesions.
- Automated Guardrails: Returns structured etiologies with confidence ratings (
HIGH / MODERATE / LOW), botanical remedies (5% NSKE / cold-pressed neem oil), and safety disclaimers before chemical intervention.
-
Format Enforcement: Gemini's
response_mime_type: "application/json"forces strict schema compliance without conversational filler or Markdown wrappers. -
Low-Temperature Determinism: Configured at
$T=0.2$ to suppress creative hallucinations and enforce adherence to retrieved scientific context. -
Explicit Action Timelines: Forces output into three actionable categories:
-
[Immediate Step - Next 24 Hours](e.g., withhold fertilizer before heavy rain) -
[Preventative Measure - Next 48-72 Hours](e.g., install yellow sticky traps) -
[Field Monitoring - Daily Inspection](e.g., scout 20 random plants for bollworm eggs)
-
- Layered Safety & Fallback Net: Prompt guardrails direct the model to ground answers in provided ICAR evidence and refuse unverified chemical dosages. In the event of API timeouts or network errors, the orchestrator triggers a deterministic, rule-based fallback safety net for common agronomic interventions.
- Colloquial Rural Dialects: Prompts instruct the model to use natural Devanagari Hindi (e.g., "पोषक तत्वों का छिड़काव" rather than dense academic jargon).
- Audio Synthesis: Native Web Speech API integration provides instant audio read-aloud in Hindi (
hi-IN) and Indian English (en-IN) for low-literacy farmers.
-
Interactive Perturbation Sandbox: Farmers slide controls to perturb microclimate values:
- Temperature:
$-5^\circ\text{C}$ to$+5^\circ\text{C}$ - Rainfall:
$-40\text{mm}$ to$+50\text{mm}$
- Temperature:
- Dynamic Counterfactual Reasoning: The agent pipeline re-evaluates the biological risk model and generates forward-looking resilience plans for impending climate shocks.
sequenceDiagram
autonumber
actor Farmer as Farmer / Client
participant Orch as FastAPI Orchestrator
participant Tools as Open-Meteo & data.gov.in
participant Risk as Deterministic Risk Engine
participant VectorDB as Neon pgvector HNSW
participant Gemini as Gemini 3.5 Flash Lite
Farmer->>Orch: POST /api/advisory (query, crop, sowing_date, GPS)
par Parallel Data Ingestion
Orch->>Tools: Fetch live satellite weather & APMC mandi rates
Tools-->>Orch: Real-time temp, humidity, precipitation, modal prices
end
Orch->>Risk: Compute GDD, moisture stress & disease favorability
Risk-->>Orch: Deterministic Risk Score (0-100) & triggered rules
Orch->>VectorDB: Cosine similarity query on ICAR research embeddings
VectorDB-->>Orch: Top-K scientific bulletin chunks & page citations
Orch->>Gemini: Synthesize grounded action plan (T=0.2, JSON schema)
Gemini-->>Orch: Structured bilingual response with 24h/48h timelines
Orch-->>Farmer: Complete Advisory Payload + Millisecond Latency Audit
| Layer | Component | Technical Role |
|---|---|---|
| Generative AI | Google Gemini 3.5 Flash Lite | Fast multimodal reasoning, vision diagnostics, and JSON-constrained synthesis |
| Vector Database | Neon Cloud Serverless PostgreSQL | Cloud Postgres with pgvector HNSW indexing over curated ICAR reference corpus |
| Embeddings | 384-Dim Hashed Vectorizer | Fast CPU-native multi-pass hashing vectorizer ( |
| Backend Framework | FastAPI (Python 3.13) + Uvicorn | Asynchronous ASGI API server running sequential multi-step agent orchestration and audit logging |
| Frontend Framework | React 19 + TypeScript + Vite 8 | Reactive single-page client with sub-second HMR |
| UI & Styling | Tailwind CSS + Lucide Icons | High-contrast rural daylight palette (#FEFADC / #241909) |
| Data Visualizations | Recharts | Interactive 14-day Mandi price projections and MSP benchmark lines |
| Speech Audio | Web Speech API | Vernacular text-to-speech audio synthesis in Hindi (hi-IN) and English (en-IN) |
| Live Telemetry | Open-Meteo API | Real-time global satellite meteorological models at exact GPS coordinates |
| Market Data | data.gov.in AGMARKNET | Official Ministry of Agriculture daily APMC arrival records & price spreads |
| Criterion | Standard Required | AgroShield Live Implementation | Status |
|---|---|---|---|
| Weather Telemetry | Genuine meteorological feed | Live Open-Meteo satellite API at exact farm coordinates | 100% Live |
| Mandi Pricing | Genuine market data | Daily APMC arrivals & prices via data.gov.in AGMARKNET | 100% Live |
| AI Grounding | Layered hallucination mitigation | Deterministic GDD scoring + pgvector RAG + strict safety guardrails | Grounded |
| Execution Observability | Non-fabricated agent traces | Real millisecond latency logs measuring each step & tool | Transparent |
| Multimodal Vision | Real visual processing | Gemini Vision inspects uploaded foliage photos in real-time | Multimodal |
git clone https://github.com/ShreyanshGiri08/AgroShield.git
cd AgroShield
python -m venv venv && source venv/bin/activate # On Windows: .\venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # Configure DATABASE_URL and GEMINI_API_KEY
python backend/init_db.py # Initialize DB schema & seed ICAR embeddings
python main.py # Server running on http://127.0.0.1:8000cd frontend
npm install
npm run dev # Frontend running on http://localhost:5173- Shreyansh Giri — @ShreyanshGiri08
- Suryansh Kumar Singh — @Suryansh-cell
- Malay Giri — @malay69420