We got called "just another LLM-wrapper startup" the other month. Best comment we've received in months. It happened after I joined Kudzai Manditereza on his AI in Manufacturing podcast (13.2K subscribers on Industry40tv). We had a great conversation about what we're building at Bardin (at the time of the podcast, we were still known as "Folio"), but honestly? The real gold was in the comment section afterward.

Full conversation here: Watch on YouTube

‍

Here's a quick recap of what we covered (though the full episode is definitely worth a listen).

‍We talked about the mess industrial sales teams deal with daily: you're selling a vision sensor, and suddenly you need to know conveyor speed, whether the plant is in Munich or Arizona, temperature range, PLC compatibility, integration conflicts, and installation constraints. Now imagine figuring all that out using PDFs, email chains, and hoping Bob from Engineering isn't on vacation.

I shared my origin story (manually writing specs for automotive parts in college), we discussed the Netflix moment for AI, the 600,000 missing manufacturing jobs in the US, and why Bardin needs to feel like the consumer AI tools people love, not clunky enterprise systems.

The podcast was released. And here's where it got really interesting: Onto the juicy stuff.

As fun as the podcast was, what came after might have been even more valuable: responding to the comment section. The best feedback doesn't come wrapped in compliments. It comes from people who've seen too many "AI solutions" that don't solve actual problems.

Deep Industrial AI: Vertical Solutions Beyond Generic Models

The Value of Application-Level Knowledge (The Best Feedback)

One commenter cut right to it: "Like all LLM-wrapper startups, it seems like a solution looking for a problem. The drivers are logistics, maintainability, reputation, support, customer preference, experience, proven performance in similar applications, cost—all the stuff that isn't in datasheets."

Totally fair skepticism. And exactly right. The highest-value knowledge isn't in the spec sheets.

That's why we're not building a general-purpose chatbot. Bardin is a vertical intelligence layer trained on the structure, nuance, and recurring friction points of real-world industrial sales workflows, an industrial sales "scribe" if you will (more on that in another blog).

Yes, LLMs are a core part of our engine, but the product value is in the background knowledge graph we build: how Bardin links sales answers to specific product outcomes, distinguishes between a spec match and a field-proven workaround, and understands that when a customer asks for a "sensor for vibrating PET line," they're really asking whether it'll survive and not false-trigger.

The AI stack is table stakes. The real challenge is encoding domain context, organizational history, and the "gut feel" judgments that live in tribal knowledge today.

Capturing The Context: Knowledge That Actually Closes Deals

Component decisions evolve through messy workflows: RFQs, back-and-forth clarification, internal judgment calls, support tickets, field rep insights. Where is all that knowledge captured? Usually nowhere. It's scattered in inboxes, saved docs, project folders, and minds.

Say a sales engineer once noted in an RFQ that "Sensor X worked in a dusty meat packaging plant because of the dual-lens alignment," and support later confirmed it after installation. That's not in any datasheet, but it's already somewhere in your company. With Bardin, that interaction becomes part of your searchable institutional memory.

Encoding Tacit Knowledge for Technical Sales

Another excellent pushback: "But How Do You Capture Tacit Knowledge?" "It would be a challenge to capture that sort of tacit knowledge as data, especially since it changes from project to project."

True—if we were only relying on structured ERP records, it wouldn't work. But tacit knowledge often is written down, it's just scattered: an RFQ follow-up, an email to a customer, an internal thread, someone logging install conditions after a field issue. AI can connect those fragments and surface them in the moment. When the next project looks similar, you're not guessing or re-solving from scratch.

Every time someone uses Bardin and adds context, "customer needed this rated to -30°C and used Model Z with custom bracket", that interaction becomes usable precedent. It's a living loop that improves as people work.

Trust and Grounding: Accuracy is Non-Negotiable for Industrial AI

One listener, John, shared how an AI tool successfully pointed him to the exact section in an Omron sensor manual, but also mentioned a friend whose LLM "confidently" claimed information was in a chapter that didn't exist.

Grounding is everything. With Bardin, every answer is tied directly to the source; datasheet, implementation manual, past RFQ exchange, or validated support resolution. And for feasibility-sensitive questions, we've built in a flow to loop in a senior engineer for review. Fast answers when they're safe, human oversight when it matters.

The Bottom Line: Specialized Expertise for Industrial Sales

Our whole thesis is: the most valuable sales and application knowledge in this industry already exists, it's just stuck in silos or people's heads. Bardin doesn't replace expertise. It gives teams (many losing decades of knowledge to retirement) instant access to everything they already know, so judgment can be applied faster and smarter.

These challenging comments force us to articulate not just what we're building, but why it's fundamentally different from the noise in the AI market. We're building because industrial sales teams deserve tools that understands the difference between matching specs and actually solving problems in the field.

Thanks to Kudzai for the conversation, and thanks to everyone for those "mean comments". ;)