Skip to content

4.5 Conclusion: two projects on one foundation

The study set out to find what a design-services vendor could actually build with AI on the assets it holds, and it ends with a concrete answer: two projects — design-reuse retrieval and the offline DFM assistant — both buildable now, both free of the labelled data the field lacks. This closing section draws them together: what they share, how each can strengthen the other, and the single pilot that should come before either build.

4.5.1 What the two share

The two proposals differ in almost everything, but they rest on the same five commitments:

  • Offline and on-premises, by assumption. Both projects are proposed for confidential, NDA-bound automotive data: customer CAD and standards often may not touch a cloud at all, so each runs entirely on the firm's own hardware and nothing leaves the building. Retrieval needs only a local vision encoder with on-premises vector and metadata databases (its LLM joins in Phase 2); the assistant self-hosts an open-weight LLM (a Qwen 32B- or Mistral-class model) on a single 24–32 GB GPU workstation. Each stack is sized in its BOM (Project 1, Project 2). The compliance constraint becomes a selling point: your data never leaves our premises.
  • Customer-owned data, multi-CAD. Nothing locks the customer in. Ingestion is STEP-first so that a mixed CAD estate is table stakes, not a special case.
  • The human keeps the sign-off. Both tools are advisory. They find and explain; a qualified engineer fixes the CAD and signs the delivery. The firm's "no flaw" liability stays with people.
  • Classical stays the authority. Deterministic geometry math and written rules decide the facts; the learning parts only search, orchestrate and explain — never measure geometry or judge physics themselves.

4.5.2 How each can strengthen the other

The two are independent builds — either delivers value on its own, and neither waits on the other to ship. But once both are in place, each makes the other better.

surfaces similar past parts
+ their DFM history

every check enriches
the indexed record

Retrieval (§4.3)
find similar past parts

DFM assistant (§4.4)
check rules, explain, report

Retrieval gives the assistant precedent. Shown a new part, it can pull up the closest past designs and whatever DFM history is attached to them — which checks were run, which issues were flagged, which customer they were for — so the assistant can weigh a rule against how similar shapes were handled before, not against the rule alone. This stays inside the rule-based, design-side world: it is past design and check history, not measured production outcomes, which the firm does not hold. And every part the assistant checks writes its findings back to that record, so each pass makes the next retrieval richer. The benefit runs both ways, and is there to take once both exist.

4.5.3 One pilot first, shared by both

Neither project should be committed to blind, and neither needs its own separate pilot. Because both rest on the same foundations — the same CAD access, the same offline stack — a single short pilot de-risks both at once, and its outcome decides what to build. It runs in three steps.

  1. Prove the CAD access and the shared scripts. Explore what CATIA can do and what it licenses, and experiment with the VBA scripts both projects rely on: getting rendered images out, converting parts to STEP, and running geometry design-checks from a STEP or CATIA file. These are common ground — retrieval needs the renders and STEP; the assistant needs STEP and the checks.
  2. Stand up the stack and prototype the orchestration. Set up the hardware, bring up the LLM orchestration, and confirm a user can interact with it through stub tools — a working skeleton that returns placeholder results — before any real tool is built.
  3. Spec, estimate, decide. With the first two steps done, write the requirement specification for each individual tool and estimate the effort from what those steps revealed — then decide whether to build Project 1, Project 2, or both.

The pilot's output is a defensible plan with a confidence band, and a clear go/no-go on each project rather than a blind commitment. The hardware and licensing each project would need are set out in its bill of materials (Project 1, Project 2).