Groundlight

What it is

Groundlight is a computer vision platform where the "model" is a plain-language question. The company's site puts the idea this way: "Point a camera, type a query, get answers." A query such as "Is any bottle missing a cap?" or "Is the pipe leaking?" becomes a detector that watches a camera feed and answers yes or no. The product is aimed at 2 kinds of reader: operations and quality teams in manufacturing, retail, facilities and security who want camera-based checks without a vision project, and developers or robotics engineers who want to add vision to their own systems. The page says "no ML or computer vision expertise required". The home page lists solutions for manufacturing, robotics, retail, facilities management and security.

What you build

In Groundlight you build detectors, not apps. The company's no-code guide explains that you can "create a binary (Yes/No) query, also known as a 'detector'", and that the detector "is essentially a computer vision model or application". The first guide step is to sign up for a free account at the dashboard and open the Demo tab. All you need is a computer and a webcam. You connect the camera, give Groundlight permission to use it, and type your question in a box labeled Query. The guide asks for a clear, objective question, and its examples are "is the cat on the couch?", "is the person wearing glasses?" and "is the thumb facing up?".

Next you set a minimum confidence level. The guide recommends 0.75, or 75%, and explains that when the model is less confident than that, "the image query will be escalated to a real human to answer in real-time". It warns that a 0.5 threshold is "like a coin toss". You can set the threshold when creating the detector and change it later. Then you gather images: the guide says to get about 10-15 different images, and an auto-capture setting takes a picture every 30 seconds so you can collect variety in a couple of minutes. The guide suggests trying different scenes so the model "will get used to a variety of scenarios".

The most unusual part is the escalation design. The company explains that images are first analyzed by ML models that are automatically trained on your data. If the models are confident enough, that is the answer. If not, "the images are progressively escalated to more resource-intensive analysis methods up to real-time human review". The result, the company says, is a system that starts working right away without a dataset you have to gather and label first. It adds a plain caveat: initially the system "will operate with high latency because people need to review the image queries", and over time the ML systems learn so queries come back faster with higher confidence. The home page calls the human side "built-in 24/7 human annotation" and describes "Invisible MLOps": its ML experts and escalation infrastructure continuously audit model performance, so models are always being updated using the latest data and science. It also names automatic model tuning and support for fast edge inference, and says models learn on the fly and adapt to changing environments: its safety harness example keeps working even when the harnesses come in different colors.

The guide also shows how you see quality. In the cloud labeling view, each captured image carries a model confidence, such as an image marked "0.54" and "Yes", which the guide reads as the model being only 54% sure. After you provide ground truth labels, the dashboard shows a Projected Machine Learning Accuracy, an estimate of how the model will do on new, unseen data. The guide says this is actually Balanced Machine Learning Accuracy, that numbers like 95% show overall projected accuracy, and that bar lengths show confidence intervals. You need enough Yes and No examples before the estimate appears. The ground truth labels you give also act as examples for the cloud labelers, helping them read your instructions and handle edge cases and ambiguous images. Counting queries exist as well, but the guide says counting is "available but for paid users" and is not on the Demo page.

For developers the route is code, and it is short. The home page says "With just a few lines of code, Groundlight makes it easy to build reliable visual applications", and "As simple as an API call and a natural-language question". It offers a REST API and a Python SDK, with code on GitHub and PyPI, and says it works on development platforms and hardware such as Arduino and Raspberry Pi. For robots there is a Python SDK plus a ROS2 package. For teams that want no integration work, the Groundlight Hub is described as "a plug-and-play computer vision appliance that gives businesses of all sizes easy access to customized computer vision applications". The products menu also lists Universal Robots and Computer Vision Tools.

The typical applications named on the site are industrial work, process control, inspections, mobile robot inspections, warehouse safety and SOP monitoring. The guide adds retail analytics and detecting proper PPE use in construction, and says the platform can work with your existing surveillance cameras.

The evidence

The site publishes no prices; signing up for a Groundlight account is free, the Demo tab needs only a webcam, the guide says counting queries are for paid users, and enterprise and robotics buyers are pointed to "Schedule a demo". On data handling, the home page says Groundlight uses "industry-standard security and privacy best practices" and that any data captured, stored or processed is encrypted. The developer menu links documentation, guides, the SDK, GitHub and a system status page, and the guide promises that a real human will respond to questions.

Three customer quotes appear, each attributed by role and not by company: a Quality Manager at a Fortune 500 manufacturer credits it with changing their quality control and inspection process; the owner of a small manufacturing business says it inspects products "without the development overhead of a typical industrial solution"; and a principal at a security managed service provider says they spent 10 years looking for a way to automate inspection of camera installations and that Groundlight Hub "verifies camera health daily for a fraction of the cost of manual checks". The company also lists a CB Insights mention among the top 100 most promising AI startups of 2024 and its place among 11 startups in Khasm Labs' Elite Batch 10 Fall Program.

Who it suits

Groundlight suits a team with cameras that already exist and a concrete yes or no question about what they see, such as a cap on a bottle, a blocked path or a harness. It also suits a developer or robotics team that wants vision as an API call. The question-first approach lets a non-specialist start without a dataset, and the home page says "Groundlight models work on day 1".

The caveats are inherent. Yes or no queries are the core unit, so open-ended image understanding is outside what the pages describe. Early answers rely on human review and come back slowly at first, as the company itself says, and pricing runs through a demo rather than a public plan table.