EU and self-hosted AI detector deployment

Teams in regulated European workflows usually care less about a global accuracy headline and more about where text is processed, whether the model can be evaluated locally, and how reviewer decisions are audited.

Short answer

Humaniser is built around EU-resident operation and reproducible local evaluation. The public API is the easiest integration path; self-hosted or private-tenant deployment should be scoped around data retention, model-update cadence, throughput, and evidence requirements.

EU data boundary

Keep submitted text inside EU-resident infrastructure for hosted inference.
Store API keys and analysis records server-side in the customer product.
Store only text digests when the review record does not need full content retention.
Document whether reviewer notes, submitted text, or only verdict metadata are retained.

Self-hosted evaluation path

The repository can be run locally so teams can test detector behavior on their own sample set before choosing a hosted or private deployment. That matters because detector performance is content-domain dependent.

cd /Users/pedro/humaniser.eu
./run-local.sh --check
./run-local.sh

What to validate before enforcement

Corpus match: evaluate on your real content types, not only public benchmarks.
False positives: inspect formal, ESL, citation-heavy, and templated writing separately.
Abstention: decide what your workflow does with manual-review output.
Analysis records: confirm that stored fields satisfy audit, DPA, and incident-review needs.
Release control: pin detector versions when policy decisions depend on stable thresholds.

Private deployment questions

A private deployment discussion should start with concrete operating constraints: expected request volume, maximum latency, supported languages, whether raw text can be logged, reviewer roles, and whether model artifacts must remain inside a customer-controlled environment.

Compliance posture

The detector is designed as a decision-support system. When a verdict can affect a person, the safer workflow is to keep a human review step, expose sentence-level evidence, retain an analysis record, and document why the final decision was made.