Responsible AI. Secure Systems. Production Engineering.
AI systems in production carry real consequences. This page describes the security engineering practices and responsible AI considerations that inform how we work.
Straight answers for your security review
The questions reviewers ask first, answered plainly. Forward this section to whoever signs off on suppliers.
- Do you hold SOC 2, ISO 27001 or HIPAA certification?
- No, and we do not claim certifications we do not hold. Any certification claim in a specific engagement will be stated explicitly and verifiably, in writing.
- Is our data used to train AI models?
- Customer data is not used to train or fine-tune models without explicit written consent.
- Can an AI agent take an irreversible action on its own?
- Critical or irreversible actions go through human approval gates, and agents run with limited permissions in sandboxed environments.
- Where do credentials and secrets live?
- In environment configuration, never hardcoded in source code. Access is role-based, with least privilege by default.
- Can we see what an agent did?
- Yes. Agent actions and data access events are audit-logged, including tool calls and outputs.
- Which frameworks inform the work?
- OWASP security guidelines, NIST AI Risk Management Framework principles, and enterprise security engineering standards. If you need a specific compliance framework, raise it at the start so we can scope accordingly.
Engineering Practices
Security and responsible AI are addressed at the design stage, not added as afterthoughts before deployment.
Security by Design
Access, secrets and logging are decided at the design stage, not added before deployment.
- Authentication and authorisation on all endpoints and agent interfaces
- Role-based access control, least privilege by default
- Data access controls and isolation between tenants
- Encryption at rest and in transit
- Secrets management via environment configuration, never hardcoded in source code
- Audit logging for agent actions and data access events
AI-Specific Considerations
Guardrails between what a model says and what the system does with it.
- Model output validation and guardrails before downstream action
- Human approval gates for critical or irreversible actions
- Agent permissions and sandboxed execution environments
- Prompt injection risk assessment in input handling
- Structured output validation and schema enforcement
- Failure handling, graceful degradation rather than silent corruption
- Model and provider evaluation against task requirements before selection
Data Practices
How personal and customer data is handled, kept and separated.
- PII identification and appropriate handling at ingestion
- Data retention policies defined per project and client requirement
- Data isolation between tenants in multi-tenant systems
- Customer data is not used to train or fine-tune models without explicit written consent
Monitoring and Observability
Seeing what an agent did, and noticing when it starts doing something unusual.
- Agent observability, logging of reasoning steps, tool calls and outputs
- Output quality monitoring with alerting thresholds
- Anomaly detection for unusual agent behaviour or output patterns
- Performance and cost monitoring per model and workflow
Security requirements for your project?
Tell us your compliance context, data sensitivity and infrastructure requirements at the start, and we will scope the right approach.