Pass a customer's security review with an LLM deployment

#Brief

Enterprise pilots often stall in security review, not in engineering. Security compliance appeared in a quarter of the Forward Deployed Engineer postings we read. The customer's CISO will ask where data goes, who can see it, how long it is kept and how you would know if something went wrong. Build a deployment that has good answers, and write them down.

Bitext messages use placeholders such as {{Order Number}} and {{Customer Name}}. Fill a sample with realistic fake personal data (names, emails, phone numbers, card-like numbers) generated by a library such as Faker, so your redaction has something to catch.

#What to build

1. Infrastructure as code for one account and region: the assistant service, a secrets manager entry for the model API key, a log store and a role for the service with only the permissions it needs.

2. A redaction step that replaces emails, phone numbers, card-like numbers and names with placeholders before calling the model, and restores them in the reply only where needed. Measure what it misses on 200 generated messages.

3. Audit logs for every model call with a request id, user, timestamp, model and a hash of the redacted input. Enforce a 30-day retention in code or storage policy.

4. A data-flow diagram and a security questionnaire covering data residency, subprocessors, encryption, access control, logging, retention, incident response and model training on customer data. Link every answer to evidence in the repository.

#Deliverables

• Repository with infra/, service/, redaction/ (with tests), docs/data-flow.md, docs/security-questionnaire.md

#How it is graded

Reviewers act as the customer's security team: they check the permissions, send messages with personal data, look for it in logs and outbound requests, and challenge three questionnaire answers.

#Stretch

Add a policy-as-code check in CI that fails if a new permission or an unencrypted resource is introduced.