Industrial AI's Blind Spot: The Expert Lore Machines Miss

When veteran technician retires, the factory's soul walks out the door. That's the multi-billion dollar problem Squint aims to solve by turning unwritten lore into AI's operating system.

Devin Bhushan, founder and CEO of the company, told PYMNTS he repeatedly hit a wall while working with industrial customers before starting Squint. "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."

This is industrial AI's blind spot. While sensors predict when a bearing might fail, they don't capture the veteran's trick for restarting a motor by listening for a specific hum or tapping a panel in a precise sequence. Squint's entire premise is that you must digitize this human expertise before any intelligence can be applied.

The Lore Machines Can't Read

The knowledge Squint targets is sensory and heuristic, locked in muscle memory and experience.

The Fixes: A technician knows a machine is overheating not from a sensor alert, but from a peculiar smell or a vibration felt through the floor grating. The solution might be an undocumented sequence of power cycles or a strategic "persuasive" tap with a specialized tool.

The Shortcuts: These are the bypasses and workarounds developed over decades that keep a line running, often conflicting with official manuals but understood by the crew. They represent the delta between theoretical design and practical, gritty operation.

XOOMAR Interpretion: This know-how is the factory's immune system. It's the accumulated adaptation to reality. Bhushan's core insight is that this immune system has been biological, residing in veteran staff. Squint is attempting to make it digital by first building what it calls a context layer.

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.

This layer is bespoke for each customer, a necessity as manufacturers fiercely guard processes like Pepsi's specific recipes. Building it was initially slow. Using a large language model to process video took 10 to 14 days per customer. Squint's solution was to train a smaller, specialized 2-billion-parameter model focused solely on watching footage and mapping it to records. This pivot from a general-purpose LLM to a domain-specific visual analyzer is a critical technical detail.

From Archive to Active Agent

Once the context layer exists, Squint shifts from archivist to active participant. Its tools become agents that diagnose and prescribe.

The Lean Manufacturing Agent automates time and motion studies. Instead of a consultant with a clipboard, the AI analyzes uploaded video, classifying every second as value-added, necessary, or waste. It makes specific, frame-level recommendations a manager can accept or reject. This tool doesn't just save labor, it provides a continuous, gapless audit of efficiency that humans physically cannot perform.

Bhushan gave two other practical examples:

• A safety briefing agent turns a lone operator's narrated 150-point checklist into a conversational, prioritized safety plan.
• A diagnostic agent for forklift service company Carolina Handling turns a vague problem description into a specific diagnosis, required parts list, and technician skill match.

The claimed results are direct: customers have reported cutting scrap rates by as much as half and reducing new process training time.

The Coming Tension Over Tribal Knowledge

This move from data capture to AI augmentation creates immediate stakeholder friction. As we've seen in other sectors like supply chain tech, such as when Human-Less Cargo Planes Score $60 Million Bet on Supply Chains, automating core operational intelligence reshapes roles and power dynamics.

Management vs. Labor: For plant managers, this is a continuity and risk-mitigation dream. It turns latent human capital into tangible, transferable intellectual property. For unions and veteran technicians, it raises red flags. Is this a tool to preserve a legacy, or the first step toward deskilling and redundancy? The value of being the only person who knows the trick evaporates if the AI knows it too.

The New Hire's Dilemma: For a novice, democratized expertise is a powerful leveler. It can flatten a years-long learning curve into weeks. But it risks creating prompt-dependent technicians who follow AI instructions without understanding the underlying mechanical principles. True expertise involves knowing when to ignore the manual, a judgment call an AI trained only on past data may not capture.

XOOMAR Analysis: This tension mirrors the fundamental shift in other expert fields. Just as Google AI Transcode Turns Talk Into Action seeks to convert fluid speech into structured commands, Squint seeks to convert fluid expertise into structured procedures. The success of both depends on whether the conversion process preserves the essential, contextual nuance.

The Endgame: Instant Factories

Bhushan's long-term vision reveals the ultimate stakes. He calls it instant changeover.

"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."

Today, retooling a line is a monumental feat of physical change and, more critically, human retraining. Every machine adjustment requires a technician with specific, often undocumented, knowledge. If an AI system contains the full procedural and troubleshooting context for both Product A and Product B, the bottleneck shifts from human knowledge transfer to the speed of robotic arms and material delivery.

This is the transformation from a fixed-asset industrial model to a flexible, demand-responsive one. The constraint moves from the cognitive (what do we know how to make?) to the purely physical (what can we build?).

What to watch next: The first metric will be mean time to proficiency for new hires on complex lines using Squint-augmented training. A significant drop validates the knowledge capture thesis. The second will be changeover time for a production line shifting between two complex products. Any measurable compression there signals the move from archival system to active nervous system. Finally, watch for labor agreements that explicitly address the use of AI-captured expert knowledge, which will be the legal and cultural battleground for this entire shift.

The quiet war isn't between humans and robots on the factory floor. It's between the tacit knowledge in a veteran's hands and the struggle to bottle it before it's gone for good. Squint isn't just building AI tools, it's running a salvage operation for industrial memory.

Why This Changes Everything

• It addresses a multi-billion dollar problem by capturing the critical, undocumented expertise that leaves with retiring technicians, which is vital for industrial continuity.
• It fundamentally shifts industrial AI from only analyzing sensor data to incorporating human sensory and experiential knowledge, potentially preventing costly downtime.
• It creates a new, essential 'context layer' for AI in manufacturing, transforming how factories preserve operational intelligence and adapt over decades.