A.T.L.A.S. — the PDF library nobody opens, rebuilt as an agent
Materials, part design, tolerances and quote direction, answered in conversation and grounded in injection molding rather than the open internet.
Nobody reads the spec sheet
Every molder has a resource library. Datasheets, shrink tables, tolerance charts, a DFM guide written in 2011. The library is real work by real engineers, and it is almost never used, because the person who needs it has one specific question about one specific part and no appetite for opening nine PDFs to find the paragraph that answers it.
The friction is not the information. It is the retrieval. An engineer deciding between a glass filled nylon and a polycarbonate blend for a housing with a snap fit does not want a materials index. They want to describe the part and be told what the tradeoff costs them.
The product
PolyAtlas is a conversational engineering agent for plastics. It handles material selection, part design and DFM guidance, tolerance questions, and quote direction, in the same conversation, in the buyer’s own language.
The engine underneath is A.T.L.A.S., the Advanced Technical Library and Application System. The naming is not decoration. It states what the thing actually is: the reference library made addressable, with an application layer that knows how to reason about a part rather than retrieve a document about one.
A library answers what a material is. An agent answers whether it is right for your part.
How it was built
The interesting work in an agent like this is not the model. It is everything arranged around the model.
The corpus came first
Datasheets, internal process knowledge, tolerance standards and the DFM material that already existed, cleaned up, structured, and scoped to the processes and resins the shop actually runs. A capable model working from a good corpus outperforms a better model guessing, every single time.
Retrieval is scoped, not open
The agent answers from that corpus rather than from the open web, and the boundary between what the library supports and what requires a human is explicit rather than emergent. Grounding is what makes the answers useful. The boundary is what makes them safe to publish.
The behavior layer took the most iteration
Hedging in ranges rather than absolutes. Asking a clarifying question when a part description is too thin to answer responsibly. Escalating cleanly instead of improvising. Most of the build time went here, not into retrieval, because this is the layer a technical buyer is unconsciously testing.
How it fits a molder’s sales motion
Contract molding runs on a long, technical pre sale. Somebody at the shop spends hours a week answering the same forty questions from prospects who may never quote. It is unavoidable work, it is how trust gets built, and it does not scale.
The agent takes the first pass. It answers the forty questions properly, at two in the morning, without a form gate. When the conversation turns into a real opportunity, the shop receives a prospect who already understands the process constraints and has done real thinking about their part. The sales engineer starts three conversations further along than they used to.
Just as importantly, it never quotes. Pricing a tool is a human judgment with commercial consequences, so the agent moves toward quote direction and hands off. Knowing where an agent must stop is most of the design work.
Why domain grounding beats generic chat
A general purpose assistant will answer a plastics question. It will usually be roughly right and occasionally confidently wrong, and in manufacturing the occasional confident error is the whole problem. A wrong shrink assumption is a steel change.
PolyAtlas is grounded in a curated injection molding corpus rather than the open web, scoped to the processes and materials the shop actually runs, and built to hedge in ranges the way an experienced process engineer does. When it does not know, it says so and routes to a person. That behavior is unglamorous and it is the reason the tool can sit on a manufacturer’s site at all.
What good looks like
Agents are easy to demo and hard to evaluate. A conversation that feels impressive is not the same as one that produced a correct answer, and a technical buyer will find the edge of a shallow system in about four exchanges.
So the evaluation set is the questions the shop already answers by hand. Those questions have known good answers from people who have run the presses, which makes them a real test rather than a vibe check. The bar is not whether the agent sounds knowledgeable. It is whether a process engineer reading the transcript would put their name on it.
Where it runs today
PolyAtlas runs as the exclusive engine behind the Plastic Parts USA platform. It is present as a floating widget across the site, as a full screen console at /polyatlas/, and as the intelligence behind the site’s technical content hub. You can see the last of those in the PolyAtlas-powered plastics guide on Plastic Parts USA, a pillar guide with five deep cluster pages built around the questions buyers actually ask.
As a standalone product it is in pilot. The architecture is deliberately portable, because the pattern is not specific to plastics. Any technical seller sitting on a library nobody opens has the same problem, the same buyer behavior, and can be given the same answer. Adhesives, fasteners, gaskets, machining, coatings, surface treatment: the shape is identical, only the corpus changes.
That is the part worth stating plainly. PolyAtlas is a product, and it is also a proof that the pattern works in a domain where being wrong is expensive. If it holds up in injection molding, where a bad shrink assumption becomes a steel change, it will hold up in most places.