Back to Posts
Dev Updates
1/2/2025
4 min read

Building in Public: Week 1 - Our Tech Stack Decision

Building in Public: Week 1 - Our Tech Stack Decision

# Building in Public: Week 1 - Our Tech Stack Decision

Welcome to our first "Building in Public" update. We're sharing our development journey transparently—the wins, the challenges, and the decisions that shape AgriLink.

This week: **choosing our core technology stack.**

## The Big Decisions

**1. Satellite Imagery: SkySat + Sentinel-2**

We evaluated five satellite providers. It came down to resolution vs. cost vs. revisit frequency.

**Winner: SkySat** (0.5m resolution, daily revisits) for high-priority fields, with **Sentinel-2** (10m resolution, free, every 5 days) as the base layer for all farms.

Why this combo? SkySat gives us the detail to catch early stress. Sentinel-2 makes it affordable to monitor huge areas. Best of both worlds.

**2. AI Model: Custom Fine-Tuned Vision Transformer**

We tested pre-trained models (ResNet, EfficientNet) and decided to fine-tune a Vision Transformer (ViT) on agricultural data. Early results are promising—93% accuracy detecting stress vs. 78% with off-the-shelf models.

The catch? Training took 6 days on 40,000 labeled field images. Worth it.

**3. Sensor Integration: MQTT + REST API**

We're supporting both protocols so farmers can use existing sensors or our recommended models. LoRaWAN for long-range connectivity in remote fields.

**4. Frontend: React + Tailwind CSS**

Fast, responsive, mobile-first. Farmers check their phones in the field—our UI has to work perfectly on a 5-inch screen in bright sunlight.

## Challenges This Week

**Problem 1: Cloud Cover**

Sentinel-2 gets blocked by clouds 40% of the time in northern Europe during winter. Our solution: temporal interpolation using past data + weather forecasts to estimate field conditions when satellites can't see through.

**Problem 2: Data Pipeline Latency**

Raw satellite images take 2-4 hours to process and deliver insights. We want this under 30 minutes. Our fix: pre-processing pipelines with AWS Lambda and edge caching. Still optimizing.

**Problem 3: Sensor Calibration**

Cheap soil sensors drift over time. We're building auto-calibration algorithms that compare sensor data to satellite moisture estimates (from radar). Tricky but doable.

## What's Next (Week 2)

- Finalizing our AI alert logic (when to notify farmers vs. when to stay quiet)

- Building the first version of our SMS/email notification system

- Interviewing 5 pilot farmers to validate our dashboard mockups

## We Want Your Feedback

If you're a farmer, agronomist, or just interested: what features matter most to you? What would you check every morning? Reply to this post or email us at hello@agril.ink.

**Pre-register to get next week's update:** [Join the waitlist](#)

— The AgriLink Engineering Team