Inspiration

Conditions on a farm change hour by hour — soil dries out, temperatures spike, water tanks deplete, and localized pest infestations emerge . Farmers cannot continuously monitor every corner of a field . While autonomous farming equipment offers a solution, machines that water or navigate without rigorous verification risk wasting water, draining batteries, or getting stranded in distant field locations . AgriGuardian was created for the Nebius x NVIDIA Global AI Hackathon to demonstrate how physical AI can watch a field, recommend intelligent interventions, and remain entirely safe, explainable, and reliable

What it does

AgriGuardian is a solar-powered farm-monitoring rover designed using a simulation-first approach . It automates field oversight through a continuous 5-step monitoring loop: Sense → Understand/Decide → Check → Act → Record . Autonomous Environmental & Crop Monitoring: Simulates real-time readings for soil moisture, temperature, humidity, water-tank levels, sunlight, battery status, and crop pest infestation levels . Power & Return-to-Station Management: Dynamically calculates solar charging efficiency and calculates a return threshold based on distance from home (15% + 0.1% per meter away, bounded between 25% and 40%), ensuring the rover turns back before its battery runs flat . Intelligent AI Advisor: Consults an NVIDIA Nemotron model hosted on Nebius Token Factory to analyze complex environmental patterns (e.g., heatwaves combined with dry soil and pest pressure) and recommend bounded actions . Deterministic Safety Gate: Enforces a strict principle — "The AI proposes, the safety gate decides" . Every recommendation (rule-based or AI-generated) must pass deterministic validation before execution, restricting actions to an approved allowlist (IRRIGATE, MONITOR, ALERT_OPERATOR, RETURN_TO_STATION) . Live Dashboard & Complete Auditability: Features an interactive Streamlit web dashboard displaying real-time metrics, historical charts, environmental alerts, and downloadable monitoring logs and immutable action audit logs .

How we built it

Language & Core Architecture: Built in Python structured into modular components: sensor simulation, power management, rule-based irrigation/pest engines, action coordinator, and deterministic safety controller . AI Inference Infrastructure: Integrated with Nebius Token Factory using an OpenAI-compatible API to call NVIDIA Nemotron models (such as Nemotron 3 Super or Nemotron Nano) . The AI returns structured JSON summaries and action proposals . Web Dashboard: Developed with Streamlit and Pandas, auto-refreshing every 5 seconds to visualize environmental telemetry, rover state, pest pressure, and AI advisor assessments . Safety Isolation & Logging: Implemented a standalone safety checker that validates parameters (e.g., blocking IRRIGATE if the water tank drops below 10%) . All decisions are written to CSV audit logs, tagging AI-originated proposals with [AI] . Automated Testing: Created a comprehensive test suite covering power thresholds, irrigation logic, pest scoring, action validation, and mock-model AI guardrails for offline verification

Challenges we ran into

Eliminating LLM Direct Control Risks: Language models can hallucinate or suggest out-of-bounds actions . We resolved this by prohibiting the AI model from controlling actuators directly, enforcing a strict JSON schema, and routing all AI proposals through the exact same deterministic safety rules applied to basic system triggers . Dynamic Distance-Based Battery Safeguards: Hardcoded battery return thresholds failed when the rover travelled farther into the field . We designed a dynamic return threshold formula (15% + 0.1% per meter) so the rover turns back earlier the farther out it navigates . Network & API :I was not able to access the TOKEN factory so I had to make a demo version of how it should look like.

Accomplishments that we're proud of

ncompromising Safety Architecture: Proven ability to reject unsafe recommendations in real-time — such as preventing an AI-requested irrigation attempt when the water tank level is critically low . Transparent Auditability: Built full traceability into every decision, giving operators complete visibility into why every action was approved or blocked . Integrated Streamlit Experience: Delivered a full-featured monitoring dashboard complete with environmental alert banners, time-series charts, pest risk tracking, and one-click CSV report downloads . Simulation-First Rigor: Validated decision loops, charging cycles, and edge cases prior to physical hardware assembly .

What we learned

Physical AI requires layered control: Non-deterministic LLMs excel at high-level reasoning and pattern recognition, but deterministic, rule-based safety controllers must hold final execution authority over physical hardware . Explainability builds trust in automation: Written reasons for every action decision are essential for farm operators to trust autonomous systems . Simulation accelerates safety design: Testing low-battery edge cases and dry-tank scenarios in software prevents physical equipment damage during early development .

What's next for Agrigaurdian

Physical Hardware & Sensor Integration: Connect real soil moisture probes, weather sensors, camera feeds, and physical irrigation pump/valve actuators behind the safety gate . Crop-Specific Threshold Customization: Expand software configuration to adjust moisture, temperature, and pest thresholds tailored to specific crop varieties and regional climates . Computer Vision for Pest & Disease Detection: Upgrade crop monitoring from simulated checks to camera-based image analysis for early pest identification . Real-World Field Testing: Deploy the physical rover to test station-to-field navigation, solar charging efficiency, and AI advisory performance under actual farm conditions .