AgroLens

In development

From field to dataset, even offline.

Year: 2026
Role: Undergraduate thesis
Context: UNICENTRO · Computer Science and Agronomy

Technologies: Flutter, Angular, NestJS, FastAPI, YOLO, PostgreSQL, Docker

Overview

AgroLens is an offline-first platform for capturing, organizing, annotating, and exporting agricultural images for computer vision datasets. It was developed as an undergraduate thesis in Computer Science at UNICENTRO, in collaboration between the Computer Science and Agronomy departments.

Source code and data remain private to protect environment configuration and research material; the public case study documents the architecture and engineering decisions.

The problem

Scattered images

Field photos end up scattered across devices, folders, and generic services.

Unstable connectivity

Internet connectivity can be unstable or unavailable during collection.

Agronomic context

Property, plot, crop, phenological stage, and location must travel with every batch.

Disconnected annotation

Annotation and export usually require tools separate from the collection process.

Data flow

Field capture
Local queue
Sync
Review
Annotation
Export

Field app

Capture or select images in Flutter.
Record property, plot, crop, and phenological stage.
Coordinates from GPS or picked on the map.
Local SQLite persistence and an offline sync queue.
Explicit states for pending, uploading, completed, or failed items.

Web dashboard

Dashboard and filters over uploads and agronomic metadata.
User, catalogue, and permission management.
Bounding box editor with annotations in YOLO format.
Export of metadata, images, and training and validation datasets.

Architecture

System flow

Field app
Flutter · offline-first
Web dashboard
Angular
API
NestJS · Drizzle ORM
PostgreSQL
metadata and audit
S3 storage
images
Background worker
uploads and jobs
AI service
FastAPI · YOLO

Layer	Technologies	Responsibility
Mobile	Flutter · Dart · SQLite	Collection, metadata, offline operation, and sync
Web	Angular · TypeScript	Management, annotation, access control, and export
API	NestJS · Drizzle ORM · PostgreSQL	Authentication, business rules, audit, and jobs
Files	S3-compatible storage	Signed uploads and image storage
AI	Python · FastAPI · YOLO	Isolated inference and assisted annotation service
Operations	Docker Compose · nginx · GitHub Actions	Reproducible environments, proxy, and automated validation

Engineering decisions

Offline first

Collection does not depend on an active connection: images and metadata are stored locally and synced when the network returns.

Metadata as part of the data

Every batch carries its agronomic context from capture time, reducing unlabeled images.

Clients separated from processing

Slow operations, such as upload completion and inference, run in workers outside the HTTP API.

Security and operations

Refresh token rotation, pre-signed uploads, rate limiting, audit trail, health checks, and documented backups.

Result and participation

The prototype integrated two working clients and demonstrated the complete flow of capture, sync, management, annotation, and export. The refactor to a dedicated backend and inference integration are still in development.

Work advised by Jotair Elio Kwiatkowski Junior, Carlos Eduardo Andrade Iatskiu, and Leandro Rampim at UNICENTRO.

Requirements gathering and domain modeling.
Flutter app and Angular dashboard development.
Offline flow and image and metadata synchronization.
Permissions, annotations, and YOLO export modeling.
Evolution to a dedicated API, relational database, object storage, and containers.