On-device ML recognition hero

NDA project

On-Device ML Inference for Real-Time Product Recognition

Mobile feature that recognizes products through the camera in real time — without relying on cloud processing

Industry

Retail / Mobile Commerce

Country

USA

Platform

Mobile (iOS / Android)

Partnership model

Turnkey Product Development

Client Context

The client was building a mobile app for in-store product discovery.
The idea was simple: user points the camera → app recognizes the product → shows details, alternatives, and offers. In reality, it didn’t work that smoothly. The first version relied on cloud inference.

It worked well in demos — but failed in real stores:

  • weak internet connection
  • slow response time (2–4 seconds)
  • users dropping off before results appeared

Challenge

  • Operational Pain: Users didn’t wait for recognition results → feature adoption was low

  • Technical Limitation: Existing ML model required server-side processing

  • Business Risk: A key feature of the app wasn’t delivering value → risk of losing competitive edge

  • Product Constraint: Recognition had to feel instant to be usable in a retail environment

On-device ML recognition demo

Solution

Instead of trying to optimize backend latency, we moved inference directly onto the device.

We took the existing model and:

  • reduced its size
  • optimized it for mobile hardware
  • converted it into a format suitable for on-device execution

The model was embedded into the mobile app, allowing real-time predictions directly from the camera stream.
This removed dependency on network conditions entirely.

Team

  • 1x

    ML
    Engineer

  • 1x

    QA
    Engineer

  • 1x

    Mobile Developer

  • 1x

    Project Manager

  • 1x

    ML
    Engineer

  • 1x

    QA
    Engineer

  • 1x

    Mobile Developer

  • 1x

    Project Manager

What We Delivered

  • On-device ML model optimized for real-time recognition

  • Mobile integration with live camera input

  • Inference pipeline running locally on device

  • Performance tuning for different device types

  • Fallback logic for low-confidence predictions

Timeline & Cost

Full breakdown is available under NDA

Discovery & Planning: 3–4 weeks

Architecture & Design: 4–6 weeks

Development: 4–6 months

Total Timeline: ~6–8 months

Discovery Phase: $25K–$40K

Design & Architecture: $40K–$70K

Development: $180K–$320K

Total Investment: $300K–$450K

Get a tailored estimate
for your project

Share a few details — we'll map scope, timeline, and cost

loading

Impact & Results

Operational:

Recognition time reduced from seconds to near-instant

Technical:

Removed dependency on backend inference and network quality

Business:

Higher feature adoption and longer in-app engagement during in-store usage

Technology Stack

  • Python
  • TensorFlow
  • ONNX
  • Rust
  • Swift
  • Kotlin
  • and more...

Not sure where to start?

Get your free project estimate within 24h or book a 30-min