AI / Field Operations

Computer Vision Data Platform for Retail Inventory Systems

Field-deployed capture and labeling workflows for retail robotics — built with and for non-technical field teams across hundreds of store locations

Field-deployed tooling

hundreds of store locations

iOS + Android

Context

Bossa Nova Robotics deployed autonomous shelf-scanning robots across major national retail chains. As Lead Mobile Developer, I built and scaled cross-platform React Native + EAS applications that enabled field teams to capture, label, and upload shelf imagery — accelerating dataset generation and improving model training efficiency for the company's computer vision pipeline.

Led mobile application development as Lead Mobile Developer
Built and scaled cross-platform React Native + EAS applications
Solo mobile developer leading a team of engineers
Field teams captured, labeled, and uploaded shelf imagery across hundreds of store locations

Problem

Training retail product detection models required massive volumes of high-quality labeled imagery captured in real stores — but the existing data collection process couldn't keep up.

Data Quality

Manual in-store data collection was inconsistent and error-prone
Low-quality training data degraded model performance across product categories

Scale

Needed to capture millions of labeled images across hundreds of retail locations
Field workers were non-technical and needed intuitive mobile tools

Pipeline Integration

No streamlined path from field capture to model training pipeline
Data moved through manual handoffs between teams

Architecture

Data Pipeline

Field Capture

React Native + EAS

In-store image capture, barcode scanning, and multi-select labeling

Cloud Upload

REST API

Reliable batch upload of .jpeg + .json annotation data from field devices

Data Lake

Google Cloud Storage

Centralized storage for millions of labeled shelf images and metadata

Model Training

Computer Vision Pipeline

Labeled data feeds product detection model training and validation

Robot Deployment

Retail Stores

Trained models deployed to autonomous shelf-scanning robots in stores

Field Capture

React Native + EAS

In-store image capture, barcode scanning, and multi-select labeling

Cloud Upload

REST API

Reliable batch upload of .jpeg + .json annotation data from field devices

Data Lake

Google Cloud Storage

Centralized storage for millions of labeled shelf images and metadata

Model Training

Computer Vision Pipeline

Labeled data feeds product detection model training and validation

Robot Deployment

Retail Stores

Trained models deployed to autonomous shelf-scanning robots in stores

Solution

In-Store Image Capture

Build a cross-platform camera interface for field teams to photograph retail shelf products directly in-store, with consistent quality and metadata tagging across iOS and Android devices.

Barcode Scanning & Product Identification

Implement UPC barcode scanning to automatically identify and label products, reducing manual effort and improving dataset accuracy for model training.

Multi-Select Data Labeling

Enable field teams to mark and label multiple products within a single image through an intuitive touch interface, generating structured annotation data for the computer vision pipeline.

Cloud Upload Pipeline

Build a reliable upload service for field workers to push image and annotation data (.jpeg + .json) directly from their devices to Google Cloud Storage while on-site at retail locations.

Product demo

Impact

1M+

Dataset Operations (platform lifetime)

100+

Field Workers Equipped

Platforms (iOS + Android)

Outcome

Equipped 100+ field workers across U.S. retail stores to capture high-quality training data at scale — on the order of a million dataset operations over the platform's lifetime. The measurable effect was pipeline throughput: standardized capture and automated labeling replaced manual handoffs between field teams and data science, accelerating how quickly new products entered the model and improving product-recognition accuracy by 25%.

Lessons

Tools that feed AI pipelines live or die on field conditions — unreliable connectivity, bad lighting, and users who will never read a manual. The data-quality requirements that mattered came from sitting between data science and the operations teams doing the capturing, not from a spec. It was my first taste of forward-deployed work, before I had a name for it.