Ask a veteran factory technician how to fix a machine that’s acting up, and the answer might sound less like an engineering manual and more like folklore: if it smells like this, hit it three times on the left side and it comes back to life. That kind of knowledge never gets written down. It lives in one person’s head. When that person retires, it walks out the door with them.
Devin Bhushan is targeting something harder to mechanize than robotic arms or autonomous vehicles: the tacit expertise that keeps a factory running. As founder and CEO of Squint, he told PYMNTS he ran into the same wall repeatedly while working with industrial customers before starting the company. “There was a lot of stuff that was in people’s heads that had never been documented,” he said. “No amount of ML or applied AI could actually help them because there was no context for that AI to work on top of.”
Squint Builds a Context Layer First
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
That gap became Squint’s founding thesis: before artificial intelligence can act on a factory floor, it needs a structured record of how work gets done. Squint built what Bhushan calls a context layer, pulling in video of workers performing jobs, informal documentation and data from a factory’s enterprise resource planning (ERP) and maintenance systems. Then it connects those pieces, linking a video of a technician using a wrench to the maintenance schedule they were following.
Each customer gets its own context layer rather than a shared model, since manufacturers guard proprietary processes closely. Bhushan pointed to Pepsi as a customer with recipes specific enough that they can’t be blended into a general model.
Building that layer with a large language model was originally slow and expensive, Bhushan said, taking 10 to 14 days per customer given the volume of video involved. Squint instead trained a small, 2-billion-parameter model dedicated to one task: watching footage and mapping it to existing records.
Agents Now Diagnose What Slows a Line
With that context layer built, Squint has moved into agents that analyze work, not just document it. Its Lean Manufacturing Agent replaces the traditional stopwatch-and-clipboard process, applying Lean and Six Sigma principles to raw video instead, according to Squint’s announcement. Uploaded footage of a work cycle becomes a gapless time and motion study, classifying every second as value-added, necessary or waste and flagging fixes tied to specific frames, recommendations a plant manager then accepts, modifies or rejects, the blog post said. The agent zooms from a single station to a full line, balancing workload against target cycle time, then out again to weigh trade-offs across a plant.
Bhushan pointed to two other examples in the conversation. One agent handles pre-work safety briefings, where a lone operator works through roughly 150 checks before starting a job. Bhushan described a conversational version of the tool that turns narrated observations into a prioritized safety plan. The other agent, built for forklift service company Carolina Handling, turns an unstructured problem description into a diagnosis of what’s wrong, which parts a technician needs, and who has the right skill set. Bhushan said customers have reported cutting scrap rates by as much as half and reducing training time on new processes.
Longer term, Bhushan is aiming at what he calls an instant changeover, where a factory could shift its entire output, from making Cheetos to making Doritos, for example, as fast as raw materials allow, rather than the months or years such transitions take today. Closing that gap, he argued, is what would let manufacturing capacity flex with demand instead of lagging behind it.
Add as Preferred Source
