Governance That Fits Around Your Stack
We design and install the layer that sits between your application and your model providers — using open-source tooling, in your infrastructure, owned by you. Then we document it so a buyer or auditor can see it.
Governance arrives late
Most teams bolt compliance on after the first enterprise security questionnaire lands. By then the pipeline is in production and every change is a migration.
Prompt-level rules leak
"Never reveal PII" in a system prompt is a request, not a control. It cannot be evidenced to an auditor, and it fails silently under adversarial input.
Evidence doesn't exist
When a buyer asks what your model returned to customer #8942 last March, application logs rarely hold a defensible answer.
Output Controls, Installed
We install a scanning layer around your model calls using established open-source libraries — typically LLM Guard or Microsoft Presidio, selected to fit your language and runtime. It runs in your infrastructure. We're not in your critical path.
- ✓Works alongside OpenAI, Anthropic, Azure OpenAI, Bedrock, and self-hosted endpoints.
- ✓Tuned in monitor-only mode against your real traffic before anything is enforced.
- ✓Fail-open or fail-closed is a per-policy decision you make and own.
- ✓No data leaves your boundary — including to us.
- ✓Open-source licences. If you never speak to us again, it keeps working.
# LLM Guard — open source, runs in your stack from llm_guard import scan_output from llm_guard.output_scanners import Anonymize, Sensitive scanners = [ Anonymize(vault), Sensitive(entity_types=["US_SSN", "CREDIT_CARD"]), ] sanitized, valid, score = scan_output( scanners, prompt, model_response )
Runs in your infrastructure. No vendor in your critical path — including us. Configuration and tuning is the work; the library is free.
Controls We Implement
Each of these is a capability in the open-source tooling we deploy. We scope which ones you need, configure them against your traffic, and hand over the runbook.
PII & PHI Redaction
Names, SSNs, card numbers, medical record numbers, addresses. Pattern and classifier based, with configurable mask formats.
Secret Exfiltration Block
Catches API keys, tokens and credentials in generated code before they reach a user's clipboard or your logs.
Bias & Fairness Guardrail
Flags protected-characteristic weighting in scoring and ranking outputs, and records the adjustment for review.
Prompt Injection Defence
Screens retrieved documents and user input for instruction-override attempts before they reach the model.
Synthetic Content Marking
Applies machine-readable markers to generated output, aligned to the EU AI Act's transparency obligations.
Evidence Logging
We wire append-only logging of every interception — rule matched, action taken, model version, timestamp — into the stack you already run.
Framework Mapping
We configure logging so every enforcement event is tagged to the control it satisfies — so evidence accumulates as you go instead of being reconstructed the week before an audit.
| Framework | What We Assess & Implement | Evidence You End Up With |
|---|---|---|
| EU AI Act | Transparency obligations, synthetic-content marking, record-keeping and human-oversight logging for high-risk use | Per-interaction technical log + policy version history |
| GDPR | Data minimisation at output, purpose limitation, records of processing for AI-mediated personal data | Redaction ledger + lawful-basis tagging per pipeline |
| HIPAA | PHI suppression against the 18 identifiers, minimum-necessary scoping, access-reason capture | Disclosure log suitable for BAA reporting |
| SOC 2 | Confidentiality and processing-integrity criteria for AI-generated output | Continuous control evidence exportable to your auditor |
| ISO/IEC 42001 | AI management-system operational controls and monitoring | Policy lifecycle + incident register |
We deliver technical controls, documentation and evidence. We are not a law firm and this is not legal advice. We are not an accredited certification body and cannot certify you against any standard — your counsel determines which obligations apply and your auditor determines whether you've met them.
Engagements
Fixed scope, fixed price, agreed in writing before anything starts. No hourly billing, no retainer required, and you own every deliverable.
Questionnaire Rescue
3–5 days · A buyer sent an AI security questionnaire and the deal is stuck
- ✓ Completed questionnaire, review-ready
- ✓ Internal note flagging every weak answer
- ✓ Gap list to turn weak into strong
Governance Assessment
2–3 weeks · The full picture, mapped to the frameworks that apply
- ✓ AI system inventory (yours to keep)
- ✓ Risk classification per system
- ✓ Gap analysis across applicable frameworks
- ✓ Costed, prioritised remediation roadmap
- ✓ 90-minute findings walkthrough
Guardrail Implementation
3–6 weeks · Priced per pipeline, not per "AI stack"
- ✓ Output filtering deployed & tuned
- ✓ Monitor-only baseline before enforcing
- ✓ Evidence logging into your stack
- ✓ Runbook + handover to your team
- ✓ 30 days post-deployment tuning
EU AI Act Readiness
Role determination, risk tiering and obligation mapping against the current timeline.
Policy & Documentation Pack
Acceptable use policy, inventory, risk procedure, AI incident response, vendor questionnaire.
Fractional Compliance Officer
Ongoing review of new AI features, questionnaire response, and regulatory change monitoring.
Bands are indicative — final price is fixed in writing after scoping and doesn't move. We're a new practice, so our first engagements carry founding-client pricing in exchange for a reference. More on that →
See it run against your own pipeline
Bring one endpoint. We'll put LAVA in front of it and show you the interception log live.