BLUEPRINT · LEGAL / SUBPROCESSORS
Subprocessors
This page has not yet been reviewed by counsel and is not legal advice.
The compliance attributes in the last column are the providers' own published claims about themselves, not certifications mAiSight holds. mAiSight is not SOC 2 audited and not ISO 27001 certified.
This is the canonical answer to "where does our data go?" It is the list referenced by the Data Processing Addendum section 6, and it is the list we give 30 days' notice against before changing.
Current subprocessors
| Provider | Purpose | Data processed | Region | Provider's own compliance claims |
|---|---|---|---|---|
| Hetzner Cloud | Application hosting | All customer data, at rest and in transit | Falkenstein, Germany (EU) | GDPR DPA · ISO 27001 |
| Anthropic | Glossary and equation extraction (default model provider), and the evaluation judge | Document text sent to the API; prompts cached for one hour | United States | Zero-retention API option · SOC 2 Type II |
| Voyage AI | Embeddings for cross-document term reconciliation | Term and symbol strings, and their definitions | United States | Privacy policy. We confirm current certification status at onboarding rather than asserting it here. |
| Cloudflare | DNS, and CDN delivery for the embed widget bundle | Widget bundle and read-side request IPs, handled by Cloudflare at the edge and never retained by us | Global edge | GDPR DPA · SOC 2 Type II |
The extraction provider is yours to choose
Anthropic is the default, but the extraction model is configurable against any OpenAI-compatible endpoint. Two consequences follow, and both are worth stating plainly to a reviewer:
- Point it at a model you host — Ollama, vLLM, anything inside your own network — and there is no third-party model subprocessor in the path at all. Your document text does not leave your infrastructure.
- Point it at a different hosted provider — OpenAI, Together, Groq, or another — and that provider becomes a subprocessor of your deployment, added under the process below.
The evaluation judge used for our own quality measurement stays on Anthropic regardless of your extraction setting; it runs against fixture data, not customer prose.
Not subprocessors
These appear in our operational tooling but do not process customer data.
- GitHub — hosts mAiSight's own source code. Documentation ingested from your GitHub repository moves under your own authorization context, not ours.
- Our evaluation tooling — used internally to score extraction quality against fixture documents. No customer prose reaches it.
Adding a subprocessor
Any new third party that would process customer data triggers all three:
- A new row in the table above, with every column filled in.
- A 30-day customer notice — email to the security contact on file, plus a dated update to this page.
- A DPA addendum where the customer has one in place.
Your rights
- Ask which subprocessors have processed your specific data. We keep provenance per document and can answer it concretely.
- Request deletion across all subprocessors. Because the default model provider operates under a zero-retention arrangement, model calls leave no residue beyond our own storage, so the deletion is scoped to our EU hosting.
- Object to a new subprocessor, and terminate before the 30-day notice expires if the objection cannot be resolved.
Contact
Subprocessor questions go to codeblackwell@gmail.com. A dedicated compliance address will be published here once the operating entity is finalized.