Data security¶

Marcel raised security in every email. Cover this before the flashy demo.

What they require¶

Drawings and know-how stay in a controlled environment
Optional local vector database / inference — example, not the only way
Cybersecurity for LLM use — clear subprocessors and policies
Engineers must trust output — transparent, traceable, no silent PLM changes

Deployment (lead with on-prem / private)¶

Option	For Theegarten
On-prem / private network	Primary story — matches their ask
Dedicated VPC	Alternative if they prefer hosted isolation
Shared public SaaS	Do not lead with this

Label slides: available in pilot vs roadmap — only claim what you can deliver.

Where data lives¶

flowchart TB
    subgraph Inside["Inside customer network"]
        D[Drawings / BOM / PLM data]
        R[RapidDraft analysis]
        L[Logs + findings]
    end
    subgraph Optional["Only if allowed"]
        X[External LLM API]
    end
    D --> R --> L
    R -.-> X

Data	Default
CAD, drawings, BOM	Processed inside
Embeddings	Project-local
Findings	Linked to PLM revision
Training on their data	No — unless contract says otherwise

Answer: “Will you train on our data?”¶

Suggested wording:

No customer-specific models, embeddings, or adapters leave your project boundary. We do not reuse your data for other customers. Optional external LLMs are feature-flagged and documented.

Engineer trust (day-to-day)¶

They need	We provide
See why	Rule name + source on drawing/BOM
Stay in control	Approve / reject each finding
Repeatability	Same package → same core checks
PLM authority	We never auto-release

Provide a short annex in the pilot contract:

Who hosts (customer vs RapidDraft)
Which LLM providers, if any
Retention and deletion
Audit log per revision

Avoid empty “GDPR-compliant” — show mechanics.

Demo tip¶

Spend 5–15 min of the hour here. Marcel’s gate; mechanical colleagues will support you if IT is satisfied.

Next: Release check demo