Autonomous Multi-Agent Agronomy, Deterministic Safety Engine & Zero-Trust Supply Chain Infrastructure
Tip
📺 Official Video Walkthrough & Live Demo:
Watch the complete architectural explanation and full working platform demo on YouTube:
👉 https://youtu.be/ACrBIGvxyBA (Click to Watch)
Important
Core Architectural Philosophy: Maximum Backend Complexity, Zero Frontend Friction.
A farmer in rural Punjab or Andhra Pradesh does not manage crypto wallets, pay gas fees, navigate multi-step forms, or read dense scientific manuals. They simply select their native dialect and upload a photo of a diseased crop. Behind the scenes, an autonomous 5-agent AI swarm, a real-time hyper-local weather engine, a compiled C++ safety firewall, and an immutable Ethereum L2 ledger validate the diagnosis and return an authentic, spoken vernacular voice advisory.
- 🎥 System Explanation & Working Demo (Video Walkthrough)
- The Crisis in Modern Agriculture (The Problem)
- The AgriNexus Innovation (The Solution)
- The 5-Agent Swarm Architecture
- Real-Time Meteorological & Rain-Fastness Intelligence
- Zero-Hallucination ICAR Agronomy Vector Core
- Statutory CIB&RC Safety & C++ Mathematical Clamping
- Edge AI Model Benchmarks & Vision Diagnostic Hierarchy
- Progressive Web App (PWA) & Offline Multi-Agent Swarm
- Live Blockchain Traceability (Base Sepolia)
- 11-Language Indic Voice Matrix (Sarvam AI)
- 3D Telemetry Control Room & Farmer UI
- Full-Stack Automated Verification Matrix
- Quick Start & Setup Guide
Watch the end-to-end platform walkthrough, multi-agent swarm telemetry, deterministic C++ safety clamping, offline PWA capabilities, and live Indic voice generation in action:
- Video Link: https://youtu.be/ACrBIGvxyBA?si=jc36sZ3cFg3yFVUC
- What is demonstrated:
- Real-time foliar pathology classification with on-device ONNX Runtime & Gemini Vision fallback.
- Live hyper-local weather & AQI spray interlock evaluation.
- Zero-hallucination ICAR/CIB&RC chemical lookup and sub-millisecond C++ dosage clamping.
- Base Sepolia blockchain cryptographic crop passport generation.
- 11-regional-language human acoustic speech synthesis via Sarvam AI Bulbul:v3 with zero-truncation audio.
- Interactive 3D laser telemetry node bus.
Smallholder farmers manage over 80% of farmland in developing nations, yet face systemic roadblocks that cause over $29 Billion in annual crop losses:
- Diagnostic Latency & Rapid Pathogen Spread:
Plant diseases like Late Blight, Rust, and Powdery Mildew spread exponentially through fields. Physical agricultural extension officers are scarce, meaning farmers often receive diagnosis weeks too late when 40% to 70% of crop yield is already destroyed. - Toxic Chemical Overdose & Soil Degradation:
Lacking exact dosage formulas, farmers rely on guesswork or local dealer advice, often spraying 2x–5x the necessary chemical concentrations. This poisons groundwater, depletes soil biology, and creates pesticide-resistant super-pathogens. - Weather Ignorance & Chemical Wash-off:
Spraying pesticides right before rain or in high winds wastes expensive chemicals, pollutes rivers, and causes toxic foliar scorching under extreme heat. - AI Hallucinations in Agronomy:
Generic consumer AI models (like raw ChatGPT) hallucinate chemical recommendations, frequently suggesting banned, lethal chemicals (e.g. Endosulfan, Monocrotophos) that violate national statutory safety limits. - Export Rejections & Lack of Traceability:
Agricultural exports worth billions (Basmati Rice, Spices, Grapes) are routinely rejected at international ports (EU / US FDA) due to untraceable chemical residue exceeding Maximum Residue Limits (MRL). - Linguistic Isolation:
Over 60% of rural farmers cannot read English technical advisory sheets or chemical safety labels, cutting them off from modern scientific research.
AgriNexus delivers a zero-trust, commercial-grade autonomous pipeline engineered specifically for real-world farming constraints:
- Tier-1 Edge Computer Vision Invariant: Always executes the trained ONNX EfficientNet-B4 neural backbone as primary vision pathology classifier (~80ms CPU inference). Gemini Vision functions strictly as a Tier-2 safety net for out-of-domain plants.
- Crop Domain Gatekeeper: Actively screens against non-target species (e.g. houseplants like Areca palm, snake plants) with a zero-chemical safety interlock.
- Hyper-Local Meteorological Intelligence: Automatically reads embedded Photo EXIF GPS or live device coordinates to fetch live temperature, humidity, wind drift, and 6-hour rain-fastness forecasts.
- Zero-Hallucination ICAR Knowledge Core: Grounded in a comprehensive agronomic vector database of 38 certified ICAR protocols, prescribing exact active chemical formulations, acre dosages, and water dilution ratios.
- Deterministic C++ Safety Interlocks: A native compiled C++17 firewall (
pybind11) intercepts all proposals, validates against statutory CIB&RC Gazette Banned Lists, and mathematically bounds dosages based on field humidity. - Immutable Ethereum L2 Crop Passports: Every verified treatment is cryptographically hashed and minted to Coinbase's Base Sepolia Layer-2 blockchain, providing tamper-proof provenance for insurance claims and international export certification.
- Acoustic Voice Hierarchy: Prioritizes authentic Sarvam AI Bulbul:v3 neural acoustic speech in 11 Indian languages online, seamlessly falling back to native on-device Web Speech synthesis when offline.
- Offline-First PWA: Installable Progressive Web App with W3C maskable vector branding, service worker asset caching (
agrinexus-offline-v3), and in-browser fallback multi-agent swarm.
flowchart TD
subgraph IN ["🌾 Field Ingestion & Meteorological Layer"]
A["📸 Leaf Image Capture"] --> B["📍 3-Tier GPS Resolution (EXIF -> Phone GPS -> Regional)"]
B --> C["⛅ Hyper-Local Weather Feed (Open-Meteo Live API)"]
A --> D["🌐 Vernacular Dialect Selection (1 of 11 Indic Languages)"]
end
subgraph SWARM ["🤖 Autonomous 5-Agent Multi-Agent Swarm (MAS)"]
direction TB
E["🧠 AGENT 1: Vision Pathology\n(Offline ONNX EfficientNet-B4 · 80ms CPU · Fallback: Gemini Vision)"]
--> F["📚 AGENT 2: Grounded ICAR RAG\n(38 Official ICAR Protocols · Zero Fallback Guessing)"]
--> G["🛡️ AGENT 3: Deterministic C++ Safety Firewall\n(CIB&RC Banned Gazette · Rain/Wind Spray Interlock)"]
--> H["⛓️ AGENT 4: Web3 Crop Passport Relayer\n(Base Sepolia L2 Gasless Smart Contract)"]
--> I["🎙️ AGENT 5: Vernacular Voice Supervisor\n(Sarvam AI Bulbul:v3 Indic Acoustic Engine)"]
end
subgraph OUT ["⚡ Dual Coordinated Outputs"]
I --> J["🔊 Weather-Aware Vernacular Spoken Advisory (Native Script)"]
H --> K["📊 3D Control Room Telemetry with Live BaseScan Links"]
end
IN --> SWARM
SWARM --> OUT
In precision agriculture, chemical efficacy is 100% weather-dependent. AgriNexus resolves field weather via a Smart 3-Tier Hierarchy:
[ 📸 Image Uploaded ]
│
▼
┌──────────────────────────────────────┐
│ Does the photo have EXIF GPS tags? │
└──────────────────┬───────────────────┘
│
┌───────────────┴───────────────┐
▼ YES ▼ NO (Web/PC Image)
[ 📍 Extract Photo's Farm GPS ] [ 📍 Capture Live Phone/Device GPS ]
│ │
└───────────────┬───────────────┘
│
▼
[ ⛅ Fetch Live & 6-Hour Forecast via Open-Meteo ]
[ 🛡️ Evaluate Rain-Fastness, Wind Drift & Spray Timing ]
-
Rain-Fastness Interlock: If rain probability exceeds
$35%$ in the next 6 hours, spraying is flagged unsafe to prevent costly chemical wash-off. Spoken voice audio explicitly commands the farmer to delay spraying until rain clears. -
Wind Drift Protection: If wind speed exceeds
$15 \text{ km/h}$ , the system alerts against chemical drift into neighboring water bodies and advises postponing until wind calms. -
Extreme Heat Safety: Under temperatures
$>36^\circ\text{C}$ , the system mandates spraying strictly at dawn (before 8 AM) or dusk (after 6 PM) to avoid rapid droplet evaporation and foliar scorching. - Active UI & Spoken Voice Feedback: Adverse weather conditions immediately render an amber Weather Alert Banner in the Farmer UI and inject factual numerical warnings directly into the 11-language Sarvam AI voice output.
AgriNexus eliminates generic AI guessing by utilizing an embedded agronomic database of 38 verified ICAR & CIB&RC research protocols covering all PlantVillage pathologies:
| Crop & Pathology | Certified Active Chemical | Certified Acre Dosage | Water Dilution | Authoritative ICAR Institute |
|---|---|---|---|---|
| Tomato Late Blight | Azoxystrobin 18.2% + Difenoconazole 11.4% SC |
150.0 ml/acre |
200 Liters | ICAR-IIVR, Varanasi |
| Tomato Early Blight | Chlorothalonil 75% WP |
200.0 g/acre |
200 Liters | ICAR-IARI, New Delhi |
| Tomato Target Spot | Pyraclostrobin 20% WG |
120.0 g/acre |
200 Liters | ICAR-NIBSM, Raipur |
| Tomato Leaf Mold | Copper Oxychloride 50% WP |
250.0 g/acre |
200 Liters | ICAR-CISH, Lucknow |
| Tomato Bacterial Spot | Streptocycline 90:10 + Copper Hydroxide 77% WP |
100.0 g/acre |
200 Liters | ICAR-IIHR, Bengaluru |
| Apple Scab | Difenoconazole 25% EC |
120.0 ml/acre |
300 Liters | ICAR-CITH, Srinagar |
| Corn Common Rust | Propiconazole 25% EC |
150.0 ml/acre |
200 Liters | ICAR-IIMR, Ludhiana |
| Grape Black Rot | Mancozeb 75% WP |
200.0 g/acre |
250 Liters | ICAR-NRCG, Pune |
| Squash Powdery Mildew | Wettable Sulfur 80% WDG |
250.0 g/acre |
200 Liters | ICAR-IIVR, Varanasi |
| Healthy Crop | Trichoderma viride 1.5% WP (Bio-Protectant) |
250.0 g/acre |
200 Liters | ICAR-IARI, New Delhi |
Before any chemical prescription reaches the farmer, it must pass the deterministic C++17 Safety Firewall (pybind11):
- Statutory Gazette Banned List Check: Validates proposals against the official Indian Insecticides Act banned schedule (Endosulfan, Monocrotophos, Dicofol, Methomyl, Carbofuran, Phorate, Triazophos, Paraquat, Methyl Parathion, Diazinon, Alachlor, etc.). Banned chemicals trigger an immediate CRITICAL STATUTORY VIOLATION lock.
-
Mathematical Humidity Attenuation:
High relative humidity (
$>80%$ ) accelerates chemical uptake and elevates foliar scorch risk. The C++ engine mathematically attenuates dosage:$$\text{Safe Dosage } (D) = \min \left( D_{\text{RAG}}, 350.0 \text{ ml/g} \right) \times \left(1.0 - \max\left(0, \frac{H - 80}{100}\right)\right)$$
AgriNexus utilizes an optimized EfficientNet-B4 deep convolutional architecture trained on 50,000+ PlantVillage images across 38 crop disease classes, exported to ONNX Runtime for high-speed edge CPU execution:
| Metric | PyTorch Base (FP32) | ONNX Runtime (AgriNexus Edge) |
|---|---|---|
| Top-1 Accuracy | 97.42% |
97.38% |
| Top-5 Accuracy | 99.64% |
99.61% |
| Macro F1-Score | 0.971 |
0.970 |
| Macro ROC-AUC | 0.993 |
0.993 |
| CPU Inference Latency | 340 ms |
82 ms (4.1x Faster) |
| Model Size | 382 MB |
75 MB (Quantized/Optimized) |
| Offline Edge Support | ❌ Requires PyTorch + GPU | ✅ Pure Numpy + ONNX CPU Engine |
- Tier 1 — Trained Neural Model (Always Primary):
- Whether online or offline, the local/edge ONNX EfficientNet-B4 model (
agrinexus_vision.onnx) is always the primary diagnostic engine. - Cloud Gemini Vision is completely bypassed during standard agronomic diagnoses, preserving the integrity and sovereignty of our custom-trained neural model.
- Whether online or offline, the local/edge ONNX EfficientNet-B4 model (
- Tier 2 — Secondary Cloud Fallback (Edge Cases Only):
- Gemini Vision is invoked only if the trained model encounters an unrecognized sample, very low confidence (<60%), or missing binary weights.
- Crop Domain Gatekeeper:
- Detects out-of-distribution non-target foliage (e.g., Areca Palm, Snake Plant, Golden Pothos) and trips an immediate Zero-Chemical Safety Interlock, informing the user that non-agricultural houseplants require zero agrochemicals.
AgriNexus is engineered as a fully installable, native-feel Progressive Web App (PWA) across Android, iOS, Windows, and macOS:
- W3C Maskable Vector Brand Identity: Equipped with high-definition 512×512 and 192×192 maskable icons (
icon-512.png,icon-192.png) and Apple Touch icons (apple-touch-icon.png) with 80% safe-zone padding to avoid Android squircle/circular clipping. - Network-First Navigation Caching: Service worker (
agrinexus-offline-v3) employs Network-First caching for HTML navigation, ensuring fresh deployments load immediately without blank screens while falling back to offline cache during outages. - In-Browser Autonomous 5-Agent Swarm: If the farmer loses connectivity in remote fields, client-side agents in
frontend/src/services/execute the diagnostic pipeline, safety checks, store-and-forward queueing, and on-device vernacular voice guidance entirely inside the browser. - Silent Standalone Mode: Automatically prompts desktop/mobile browser visitors to "Add to Home Screen", while running silently in native fullscreen when launched from the home screen.
Every crop diagnosis is permanently minted to Coinbase's Base Sepolia Ethereum Layer-2 blockchain:
| Property | On-Chain Detail |
|---|---|
| Smart Contract | CropPassport.sol (Solidity 0.8.20 + OpenZeppelin) |
| Contract Address | 0xDd819A09aff9A62D1F6Ad662c6cC34d4B5D7DAd7 |
| Network | Base Sepolia Testnet (Chain ID: 84532) |
| Explorer | 👉 View Live Transactions on BaseScan |
struct PassportRecord {
uint256 timestamp; // Block timestamp of verified diagnosis
string imageHash; // SHA-256 fingerprint of the farmer's leaf photo
string diagnosis; // Pathology classification (e.g. "Tomato Late Blight")
string treatmentHash; // Cryptographic hash of prescribed ICAR treatment
bool isSafe; // Verified safe flag from C++ safety core
}AgriNexus natively speaks and writes in 11 Indian Regional Languages with a deterministic priority hierarchy:
- Online Primary (Sarvam AI Bulbul:v3): High-fidelity neural voice synthesis (
shubhspeaker) executing whenever internet connectivity is detected. Includes interactive audio replay controls (🔊 सुनो). - Automated Quota Failover & Offline Fallback (Native On-Device Web Speech): If Sarvam AI encounters daily request quota limits (HTTP 429/402/403) or the device is offline (
!navigator.onLine), AgriNexus automatically fails over to the mobile device's built-in Web Speech API (window.speechSynthesis) across all 11 Indian languages with zero audio interruptions.
| Language | Native Script | Agronomic Greeting | Language | Native Script | Agronomic Greeting |
|---|---|---|---|---|---|
| Hindi | हिन्दी | किसान भाई | Bengali | বাংলা | কৃষক ভাইয়েরা |
| Punjabi | ਪੰਜਾਬੀ | ਕਿਸਾਨ ਵੀਰੋ | Marathi | मराठी | शेतकरी मित्रांनो |
| Telugu | తెలుగు | రైతు సోదరులారా | Gujarati | ગુજરાતી | ખેડૂત મિત્રો |
| Tamil | தமிழ் | விவசாய சகோதரர்களே | Odia | ଓଡ଼ିଆ | କୃଷକ ଭାଇମାନେ |
| Malayalam | മലയാളം | കർഷക സുഹൃത്തുക്കളെ | English | English | Dear Farmer |
| Kannada | ಕನ್ನಡ | ರೈತ ಮಿತ್ರರೇ |
- Farmer Field Interface (
FarmerView.jsx): Mobile-responsive, single-tap photo upload (dual Camera & Gallery support), live Farm Weather HUD Badge (⛅ 34°C · 55% Humidity · Safe to Spray ✓), bilingual regional language selector, and an automated Sarvam AI audio player with manual replay triggers. - Swarm Control Room (
TelemetryView.jsx): Real-time 3D cybernetic canvas featuring progressive 850ms laser line propagation, dynamic bot orbital wobbles, shield rotations, harmonic sound ripples, and a live cryptographic ledger with direct[BaseScan ↗]one-click explorer links.
Every pull request and build is verified through a 51-test automated matrix across 3 distinct technology runtimes:
| Test Suite | Framework | Scope | Passing Tests |
|---|---|---|---|
| Deterministic Core & Backend | Pytest (pytest tests) |
C++ safety engine, ICAR RAG retrieval, KVK Haversine resolver, weather spray interlocks, and crop domain firewall | 28 / 28 ✅ |
| Client Swarm & UI Components | Vitest (vitest run) |
Edge multi-agent swarm, meteorological voice gates, offline store-and-forward queue, FarmerView HUD, and logo branding | 18 / 18 ✅ |
| Smart Contract Provenance | Hardhat (npx hardhat test) |
CropPassport Solidity contract, Base Sepolia event emissions, owner-only authorization, and record integrity | 5 / 5 ✅ |
| Production Bundle Compilation | Vite (vite build) |
TypeScript/JSX compilation, asset hashing, Tailwind CSS tree-shaking, and PWA manifest linking | 0 Errors ✅ |
Backend (backend/.env):
SARVAM_API_KEY=your_sarvam_api_key_here
GOOGLE_API_KEY=your_gemini_api_key_here
BASE_SEPOLIA_RPC_URL=https://sepolia.base.org
DEVELOPER_PRIVATE_KEY=your_wallet_private_key_here
CROP_PASSPORT_CONTRACT_ADDRESS=0xDd819A09aff9A62D1F6Ad662c6cC34d4B5D7DAd7
# Databases & Storage (New)
MONGODB_URI=mongodb+srv://<user>:<password>@cluster0...
CLOUDINARY_CLOUD_NAME=your_cloud_name
CLOUDINARY_API_KEY=your_api_key
CLOUDINARY_API_SECRET=your_api_secret
PORT=8000
HOST=0.0.0.0Frontend (frontend/.env):
VITE_SARVAM_API_KEY=your_sarvam_api_key_herecd backend
python -m venv venv
# Windows: venv\Scripts\activate | Linux/macOS: source venv/bin/activate
pip install -r requirements.txt
python main.pyBackend runs on http://localhost:8000 with WebSocket telemetry at ws://localhost:8000/ws/telemetry.
cd ../frontend
npm install
npm run devFrontend runs on http://localhost:5173.
# Backend Pytest Suite
cd backend && python -m pytest tests
# Frontend Vitest Suite
cd ../frontend && npm test -- --run
# Smart Contract Hardhat Suite
cd ../contracts && npx hardhat test- Admin Dashboard & Telemetry: Full visibility into platform usage, global scan history, and role-based user management via a secure React Admin dashboard backed by MongoDB Atlas.
- Professional PDF Generation (jsPDF): Farmers and KVK extension workers can generate highly detailed, dynamically styled PDF diagnostic reports featuring:
- Embedded Cloudinary crop images
- Red/Green color-coded Safety Spray Status
- Exact ICAR generic chemical prescriptions
- Base Sepolia Blockchain verification links
- Nearest KVK location referrals for local assistance.
AgriNexus is built primarily for B2G (Business-to-Government) deployment. By acting as an unbiased, ICAR-compliant oracle, it serves as a critical infrastructure layer for state agriculture departments and KVKs.
To maintain our zero-hallucination, unbiased agronomic integrity, AgriNexus avoids direct corporate pesticide bias. Instead, it monetizes via a Generic Diagnosis + Sponsored Fulfillment model:
- The AI neutral-diagnoses the required generic chemical (e.g., Copper Oxychloride 50% WP).
- Regional verified sellers optionally sponsor the local fulfillment marketplace.
- Furthermore, aggregate immutable disease heatmaps are monetized for macro-level crop insurance underwriting.
Distributed under the MIT License.