Product · UW × Microsoft × Nelson Farms · 2025

FarmGazer

Designed, built, and shipped FarmGazer, an AI-powered crop monitoring system combining pan-tilt cameras, microclimate sensors, and multimodal LLMs into a single mobile dashboard. Deployed on Nelson Farms, a 7,500+ acre working farm, sponsored by Microsoft and UW's Global Innovation Exchange.

I owned product definition, end-to-end user experience, industrial enclosure, and hardware development. The core product pivot I helped drive: instead of solving single-issue detection, we reframed FarmGazer as holistic crop monitoring that keeps farmers in the loop rather than replacing them.

Role

Product Builder
Definition · UX · Enclosure · Hardware

Partnership

UW GIX
Microsoft × Nelson Farms

Type

Working Prototype
Field-Tested · 7,500+ Acres

Stack

GPT-4o mini · Raspberry Pi
Azure SQL · Mobile App

FarmGazer hero: farmer holding phone showing the FarmGazer app, pan-tilt camera mounted in the field

FarmGazer · Deployed on Nelson Farms (mobile dashboard + pan-tilt camera)

Watch full FarmGazer demo

Monitoring crops manually is challenging for large farm owners (1,000+ acres) due to a declining workforce, yet timely detection of issues is crucial for minimizing economic losses. FarmGazer moves farmers beyond single-issue detection into holistic, LLM-powered monitoring that scales across thousands of acres.

Why FarmGazer matters: 7% of large farms control 73% of farmland; labor costs surged 32% since 2017 while employment declined 15%; 14% of yield losses could be prevented

Why FarmGazer matters · The numbers behind the problem

Our approach: reframing the problem through field research and stakeholder conversations with farmers

Approach · Reframe the problem + identify knowledge gaps through user research

What I owned

Product definition: co-scoped FarmGazer's core premise, pushing the team from single-issue detection toward holistic LLM-powered monitoring, and translated user research into shippable product scope.
User experience flow: defined the three-tab information architecture (Overview, Detail, Copilot) that turns raw AI output into prioritized, actionable alerts.
Mobile dashboard: designed visual alerts with traceable AI reasoning so farmers see the why, not just the what.
Override controls: gave farmers the ability to edit AI-generated tags and reprioritize alerts, keeping humans in the loop.
Industrial enclosure: designed and fabricated the pan-tilt camera housing with waterproof roof and solar panel for autonomous field operation.
Hardware integration: assembled the Raspberry Pi, Witty Pi, BME 280 sensor, Pi Camera, and servo motors into a field-ready working prototype.

FarmGazer hardware: Witty Pi, Raspberry Pi, BME 280 sensor, Pi Camera, servo motors, and the deployed pan-tilt camera unit in the field

Hardware + Enclosure · What I designed and built, deployed on-site

Data processing pipeline: Sensor Data and camera imagery flow through Azure automation into GPT-4o mini Vision LLM, then into Azure SQL Database with image_id, temperature, humidity, AI_analysis fields

Data Pipeline · How sensor data and imagery flow into LLM analysis

Three AI features I defined

AI Detection: daily automated capture flags abnormal crop conditions and pushes prioritized alerts, ending the need for constant on-site inspections.
AI Analysis: each detection is paired with microclimate context and a plain-language explanation (Problem Category, Analysis Summary, Interpretation) to build farmer trust in the result.
AI Insight: a natural-language copilot lets farmers query historical conditions on their own farm, backed by their own data and an agricultural knowledge base.

FarmGazer mobile features: four screens showing pan-tilt hardware, daily AI-tagged monitoring, detailed microclimate insights, and historical copilot

Mobile App · Three AI features shipped to farmers

What we learned about trust between users and AI

Users felt uncomfortable when AI produced lengthy but incorrect analyses. Design pivot: users can edit AI-generated tags at any time. Principle: AI should support users, not make decisions for them.
Users worried about giving their farm data away to train commercial AI models. Design pivot: keep AI analysis concise, and question whether inference could stay within edge models rather than the cloud.
Trust comes from transparency and reversibility, not raw model accuracy.