FieldGuard

Inspiration

India extracts more groundwater than China and the United States combined—over 250 cubic kilometers every year—and 90% goes directly to agriculture. Yet across smallholder farms, irrigation is still run almost entirely on guesswork: pumps are switched on until fields are submerged under unmetered flood irrigation. This wastes up to 60% of pumped water through solar evaporation and deep percolation, exhausts local aquifers, and consumes more than 20% of rural electric grid capacity.

When we spoke with farmers and rural agronomists, we realized conventional agtech solutions fail not because the science is missing, but because the interface is broken. A farmer with wet, mud-stained hands standing in a field in Purba Bardhaman or Nashik cannot navigate complex English dashboards, decode multi-tab soil metric tables, or interpret raw moisture charts. They don't want raw sensor telemetry; they need a direct, timely answer in their own language: "Irrigate 4,200 liters at 6:30 AM before peak solar evaporation, and keep the onion lateral valves shut."

We built FieldGuard to turn rigorous biophysical agronomy into an intuitive, hands-free conversational voice experience powered by Alexa+, Amazon Bedrock, and the Model Context Protocol (MCP).

What It Does

FieldGuard is an autonomous climate-adaptive precision irrigation engine that acts as a 24/7 agricultural advisor in the farmer's pocket:

1. Daily FAO-56 Penman-Monteith Evapotranspiration: Computes hyper-local reference crop evapotranspiration \(ET_0\) from live solar radiation, wind speed, relative humidity, and temperature. Using crop-specific growth coefficients \(K_c\), it calculates exact daily crop water needs:

$$ ET_c = K_c \times ET_0 $$

1. Dynamic Root-Zone Depletion Ledger: Tracks root-zone soil water balance using daily soil moisture depletion accounting:

$$ D_{r, i} = D_{r, i-1} - (P_i - RO_i) - I_i - CR_i + ET_{c, i} + DP_i $$

It calculates Readily Available Water \(RAW\) based on soil texture Available Water Capacity \(AWC\) and root depth \(Z_r\), warning the farmer days before crops reach cellular drought stress (\(D_r \ge RAW\)):

$$ RAW = p \times TAW = p \times [1000 \times (\theta_{FC} - \theta_{WP}) \times Z_r] $$

1. Conversational Voice Cockpit (Alexa+): A full-duplex voice interface built on the 2025-11-25 Model Context Protocol specification. Farmers speak naturally in Hindi, Bengali, or English to check tank levels, schedule water budgets, switch irrigation methods, and log field events.
2. Sub-5ms Client-Side Voice Barge-In: Real-time voice interruption allows farmers to interject mid-sentence (e.g. saying "Alexa, stop" or "Ruko"), instantly halting speech synthesis without waiting for server roundtrips.
3. Multi-Parcel Water Budgeting: When water is scarce (e.g. only 800 Liters remaining in the tank), FieldGuard prioritizes water allocation strictly by root moisture urgency, instructing the farmer exactly which lateral valves to open and which to keep closed.
4. Quantified Environmental Ledger: Translates deferred irrigation volume into auditable energy and carbon offset metrics:

$$ \text{Energy Saved (kWh)} = \left(\frac{V_{\text{conserved}}}{\text{Pump Flow Rate}}\right) \times P_{\text{pump}} \times 0.82\text{ kg CO}_2/\text{kWh} $$

1. Zero-Reading Specimen Cards: Designed for rural accessibility with photographic specimen cards for regional soil textures and crops, minimizing text dependency.

How We Built It

We engineered FieldGuard on a four-tier serverless cloud architecture built end-to-end on AWS:

• Frontend & Voice Cockpit (Tier 1): Next.js 16 App Router, TypeScript, and Tailwind CSS. The audio subsystem uses the Web Speech API with dual-stream recognition and client-side barge-in interrupt handlers.
• Agentic Reasoning & Orchestration (Tier 2): Amazon Bedrock (Converse API, us.amazon.nova-micro-v1:0 / Claude 3.5 Sonnet) deployed in region ap-south-1. It handles multi-turn intent extraction, parameter resolution, and mathematical reasoning while strictly enforcing agronomic guardrails.
• Self-Hosted MCP Server (Tier 3): A self-hosted Model Context Protocol server at /api/mcp strictly implementing the official 2025-11-25 MCP specification over Streamable HTTP with JSON-RPC 2.0. It exposes six precision tools:
  • get_farm_context: Live farm profiles, pump ratings, and plot metadata.
  • get_weather_context: Hyper-local solar radiation, humidity, and rainfall probability.
  • get_plot_water_status: FAO-56 root depletion \(D_r\), stress thresholds, and days until drought.
  • allocate_limited_water: Rationing constrained water budgets across competing crop blocks.
  • compare_irrigation_methods: Calculating water, electricity, and diesel savings across Drip (90% eff), Sprinkler (75%), and Surface Flood (60%).
  • update_water_reserve & log_irrigation_event: Bidirectional synchronization of tank volume and past watering events.
• AWS Serverless Data Fabric (Tier 4):
  • Amazon DynamoDB: CropPulse-Farms and CropPulse-DecisionLogs tables in ap-south-1 provide single-digit millisecond persistence of farm records, soil parameters, and ledger histories.
  • Amazon Polly: Synthesizes neural multilingual voice responses for Hindi, Bengali, and Indian English accents.
  • AWS Cognito: User Pools and Identity Pools secure farmer profiles and authentication sessions.
  • Open-Meteo: High-resolution hourly atmospheric telemetry for solar radiation \(R_s\), wind speed \(u_2\), and vapor pressure deficit \(VPD\).

Challenges We Ran Into

1. Full-Duplex Voice Latency and Barge-In: Standard browser speech synthesis queues audio utterances asynchronously. If a farmer interrupted the agent with a voice command, the browser would continue playing the previous response. We solved this by architecting a client-side audio interruption controller that halts synthesis and cancels active audio buffers within 5 milliseconds, clearing the pipeline before the user finishes speaking.
2. Self-Hosted MCP over Streamable HTTP: Implementing the 2025-11-25 Model Context Protocol specification directly in Next.js App Router route handlers required engineering Server-Sent Events (SSE) streaming and chunked transfer encoding compatible with both MCP client inspectors and browser voice agents without relying on heavyweight container sidecars.
3. Database Grounding vs. LLM Hallucinations: During multi-turn voice conversations, language models tend to parrot numbers mentioned earlier in chat history or invent round numbers like 10,000 L. We inverted the resolution priority in our tool executor to treat Amazon DynamoDB as the immutable single source of truth, enforcing strict prompt constraints where the model must query get_farm_context and quote live database attributes directly.
4. Colloquial Agricultural Multilingual Support: Indian farmers describe tank volumes using colloquial terms ("ek chauthai", "aadha", "pura khali", "dui bigha"). We built semantic normalization tables across Hindi, Bengali, and English to map conversational expressions into exact numeric liters and percentages before tool dispatch.

Accomplishments That We're Proud Of

• Verifiable Water & Energy Savings: Benchmark calculations demonstrate up to 35% water savings and 28% electricity reduction by shifting irrigation to cool morning hours and matching water volume strictly to daily crop evapotranspiration.
• Flawless Hands-Free Voice Experience: Built a true voice duplex loop where farmers never need to click mic buttons between conversational turns.
• Strict Specification Conformance: Engineered a clean, self-hosted implementation of the 2025-11-25 Model Context Protocol over Streamable HTTP with 100% test coverage across all tool schemas.
• Production-Grade AWS Infrastructure: The application runs completely serverless on AWS (Bedrock, DynamoDB, Polly, Cognito) with near-instantaneous page transitions and low operational overhead.

What We Learned

• Deep architectural patterns of the Model Context Protocol (MCP) and how tool-calling agents fundamentally outperform static rule engines in ambiguous, real-world user scenarios.
• The intricate biophysics of the FAO-56 Penman-Monteith method, including vapor pressure deficits, psychrometric constants, and crop coefficient curves across growth stages.
• Principles of designing for rural digital equity: high contrast, zero-reading visual cues, and dialect-tolerant voice recognition.

What's Next for FieldGuard

• LoRaWAN IoT Sensor Ingestion: Direct integration of low-cost capacitive soil moisture probes and water flow meters into our MCP server pipeline.
• Automated Pump Control: Connecting our decision engine to solar pump Variable Frequency Drives (VFDs) and smart relay solenoids for autonomous valve switching.
• Offline Edge Deployment: Compiling the FAO-56 engine into WebAssembly for offline edge execution on low-cost Android field tablets when cellular connectivity is intermittent.

Built With

• agricultural-technology
• alexa+
• amazon-bedrock
• amazon-dynamodb
• amazon-polly
• aws-cognito
• clean-water
• climate
• environmental-ai
• fao-56-agronomy
• json-rpc
• lucide-icons
• model-context-protocol-(mcp)
• next.js
• node.js
• open-meteo-api
• react
• rest-api
• serverless
• tailwind-css
• typescript
• voice-ai
• web-speech-api