The most expensive part of a factory’s computer vision system isn’t the camera. It’s the data scientist you have to ask, months later, to change what the camera is looking for. Groundlight AI, a Seattle startup, is betting you can replace that conversation with a line of code. Its platform lets a developer ask a camera a question in plain English, ‘Is the safety gate closed?’ or ‘Are there more than three boxes on the pallet?’, and get an answer back through an API, no custom model training required [Perplexity Sonar Pro Brief].
It’s a simple premise with a complicated backend, combining off-the-shelf deep learning, human reviewers for tricky cases, and a system that learns from feedback. The goal is to make visual intelligence as easy to query as a database, and to do it for the messy, physical world of manufacturing floors, retail stockrooms, and warehouse docks.
The natural language wedge
Groundlight’s core product is an API. A developer sends an image or video stream alongside a natural language query, and the service returns a yes, no, or count. The technical magic is in how it handles the ‘vision’ part. For common queries, it can use a pre-trained vision-language model. For novel or nuanced tasks, it can route the image to a human annotator for a quick label, using that feedback to refine an automated model for next time [Perplexity Sonar Pro Brief]. This human-in-the-loop system is the crutch that lets the platform promise robustness from day one, especially in environments where lighting, angles, and obstructions are variables, not constants.
The company is targeting verticals where the cost of failure is high but the budget for AI talent is low: industrial automation, quality control, retail analytics, and robotics [Perplexity Sonar Pro Brief]. A recent feature launch, ‘Counting Mode,’ explicitly goes after inventory and production tallying, a tedious but critical task [Yahoo Finance, May 2025].
Founders from Amazon and the factory floor
The team’s background suggests they understand both the software and the shop floor. CEO Avi Geiger was a hardware architect at Microsoft and later co-founded PicoBrew, where he built an automated brewery [The Seattle Times, retrieved 2026]. CTO Leo Dirac came from Amazon, where he led applied ML and AI teams [Madrona].
They’ve assembled a small team, estimated at under 25 employees, and are based in Seattle. The company emerged from stealth in April 2023 with a $10 million seed round led by Madrona Venture Group, with participation from Greycroft, Founders Co-op, Flying Fish, Ascend, and Essence VC [VentureBeat, April 2023].
| Co-Founder | Title | Key Background |
|---|---|---|
| Avi Geiger | CEO | Former Microsoft hardware architect; co-founder of PicoBrew (automated manufacturing) [The Seattle Times, retrieved 2026]. |
| Leo Dirac | CTO | Former Amazon AI/ML leader, with a background in applied machine learning [Madrona]. |
Where the model meets the metal
Groundlight’s go-to-market strategy hinges on compatibility with inexpensive, ubiquitous hardware. The company explicitly promotes using off-the-shelf network cameras, lowering the barrier to deployment [groundlight.ai]. For more integrated solutions, they offer the Groundlight Hub, an appliance for on-site, real-time vision processing [Groundlight AI]. This focus on the edge is pragmatic, avoiding latency and data privacy concerns of sending all video to the cloud.
The company has also released an open-source framework for building visual reasoning agents, a move that seems aimed at cultivating a developer community [MarkTechPost, March 2025]. For now, however, the service remains in a pilot phase, available to a limited set of users [VentureBeat, April 2023].
The risks in the loop
The bet is elegant, but the path to scale is lined with specific challenges. Groundlight must prove its model can work consistently without becoming dependent on expensive human review, a cost that could crater unit economics. It also operates in a crowded field.
- The accuracy-cost trade-off. The human-in-the-loop is a brilliant bootstrap, but it’s a cost center. The platform’s long-term viability requires its automated models to rapidly achieve high accuracy, minimizing the need for paid human intervention.
- Competitive landscape. Groundlight is not alone in trying to democratize computer vision. Roboflow excels at the dataset management and model training pipeline for developers who want more hands-on control. Landing AI, founded by Andrew Ng, also promotes a platform for building vision AI with less code, with a strong focus on manufacturing.
- The pilot paradox. Being in a controlled pilot allows for refinement, but it delays the true test: how the system performs when hundreds of customers, with thousands of unique queries, hit it simultaneously.
The next twelve months
The coming year will be about moving from pilot to product. Key milestones will include a full public launch, the announcement of initial paying customers, and likely a Series A fundraise to scale the team and go-to-market efforts. The company will need to demonstrate that its system can handle the diversity of real-world queries while keeping computational and human-review costs in check.
For Groundlight to succeed, it doesn’t need to beat the hyperscalers at building the most powerful vision model. It needs to beat the incumbent solution for small and mid-sized industrial firms, which is often a clipboard, a pair of human eyes, and a prayer that nothing gets missed. That’s a market measured in millions of cameras, not millions of parameters.