Computer vision case study: 3 years embedded

This computer vision case study is less about the model and more about the team an AI company decided to keep. Groundlight is a US computer vision SaaS for industrial safety. The core product is an API that lets customers ask natural-language questions about an image or video stream (is a hard hat on, is the safety guard in place, is a door open) and get an immediate answer back, with human reviewers in the loop when the model is uncertain. The product applies across manufacturing, facilities, and any environment where machine vision combined with human verification shortens time-to-detection.

As Groundlight grew from an MVP into a SaaS product, the bottleneck was engineering bandwidth on both the frontend and the backend. The customer-facing interface needed real product work (managing detectors, viewing detector improvement over time, surfacing labeler performance reports), and the in-house team did not have headcount to ship at that pace.

Groundlight found Leanware through a Google search. The brief was straightforward: extend the internal engineering team with frontend and backend developers who could be trusted to work on the production system without heavy hand-holding. For a company weighing whether to hire computer vision developers, the shape of that brief is worth noting. Groundlight builds the vision technology itself; what it needed to hire was product engineers who could work responsibly around a production ML system.

The engagement was dedicated-team capacity, retained on a monthly basis, and ran from December 2022 until Groundlight was acquired in early 2026. The embedded team was one senior full-stack engineer, three mid-level full-stack engineers, and a product designer. AI fluency was a baseline on the team rather than a specialty rate, which mattered here because Groundlight's product is a computer vision system and the engineering work routinely touched the ML feedback loop. The point of contact rotated between engineers depending on which part of the system was shipping that sprint.

The work spanned the production system. On the frontend, the team owned the customer-facing detector management UI in React Native, the interface a customer uses to create a detector, upload images, label new examples, and watch detector accuracy improve over time. On the backend, the team worked in Django and Python, integrating with the labeling teams that supply human-in-the-loop answers when the computer vision model is uncertain about a frame.

A representative thread of work the team shipped: when the algorithm produces an incorrect or low-confidence answer, the system routes the image to a labeling reviewer; the reviewer response is then incorporated into the detector training loop. Building that loop reliably (labeler queueing, reviewer feedback ingestion, accuracy-over-time visualization) required engineering across the stack and ongoing iteration with the data team.

The team also shipped image bounding-box authoring (so customers can mark regions of interest before training a detector), per-customer account metrics, signup flows, and a series of UI/UX improvements aimed at manufacturing operators rather than the engineering audience the early product was shaped for. Tech stack: React Native, Django, Python, PostgreSQL, AWS, Kubernetes, GitHub Actions.

Communication ran on Slack and Jira. Code went through the Groundlight internal review process; the team worked increasingly independently as code quality held over time.

The engagement ran three years, from December 2022 until Groundlight was acquired in early 2026. The frontend went from intermittent shipping to substantial launches every two to three weeks, and the engagement expanded from a frontend trial into full-stack ownership over multiple parts of the product. Code-quality review held to the point where the team worked increasingly independently. The Clutch review came in at 5.0 out of 5 across every category, with $120,000 invested at that point.

What makes this computer vision case study distinctive is who the client was. A company that builds computer vision for a living chose to extend its team with outside product engineers, and kept that team for three years, up to the acquisition. The qualitative thread that ran through the relationship was trust. Morgan Venable, Head of Product at Groundlight, was on record describing the engagement as "so much more" than an outsourcing partnership and noting the senior engineers were "always around to support us." Decision input flowed in both directions. Groundlight respected the technical input the team brought, and the team took ownership of code quality and direction without close management.

The same long-running embedded pattern runs through Leanware's other dedicated-team work, including ConnectCapable's AI traffic incident platform and Greyhound Engineering's industrial analytics team.