Trust
Responsible by design.
AI systems earn trust through constraints, evidence and clear ownership — not slogans. We design for oversight, privacy and security from the first workshop to the handover.
Principles
Seven commitments that shape the work.
These are the questions we ask on every engagement. They shape architecture, data flows, interfaces and the evidence we leave behind.
-
Human oversight
Every system has a named owner and a clear escalation path. People can review, override and stop automated actions that affect customers, money or rights.
-
Privacy by design
We collect the minimum data the use case needs, set retention limits and keep personal information out of prompts, logs and test sets unless it is necessary and approved.
-
Security first
Threat modeling starts with the architecture, not after launch. Access follows least privilege, secrets are managed, and dependencies are tracked.
-
Evaluated, not assumed
We define what good looks like before we build, then evaluate accuracy and refusal behavior and test for prompt injection, bias and other failure modes. A demo is not evidence.
-
Transparent and documented
Architecture decisions, data flows, model choices and known limits are written down. Clients leave with documentation they can operate and audit.
-
Vendor-neutral
We recommend the model, platform and tooling that fit the problem, the budget and the constraints. No reseller quotas, no default stack.
-
Inclusive and accessible
Interfaces are designed with WCAG 2.2 AA in mind, built for keyboard and screen-reader use, and written in plain language. AI should widen access, not narrow it.
In practice
How we put them into practice.
Controls, not promises. The same controls apply whether we are building a prototype or operating a production system.
-
Client data and model training
We do not use client data to train models without written consent. Where training is part of the work, it happens in a governed environment with a documented purpose and an exit plan.
-
Data minimization
We map data flows before we build and collect only what the use case requires. Sensitive fields are excluded, masked or tokenized where possible.
-
Least-privilege access
People and services receive the access they need for the task, for as long as they need it. Production access is limited and reviewed.
-
Secrets management
API keys, credentials and tokens live in a managed secret store — never in source code, prompts or shared documents.
-
Encryption
Data is encrypted in transit and at rest through the capabilities of our cloud providers, with managed key storage.
-
Multi-factor authentication
Accounts that touch client systems or data require multi-factor authentication. We prefer hardware keys or passkeys where the provider supports them.
-
Logging and audit trails
Agent actions, model calls and administrative changes are logged where the platform allows. Logs are kept for investigation and review, and kept free of unnecessary personal data.
-
Evaluation and red-teaming
Before launch, we test against a documented evaluation set and run adversarial checks: prompt injection, data exfiltration, unsafe tool use and refusal failures. Results are recorded, not just discussed.
-
Incident response
Every engagement has an incident path: who is called, what gets contained, how clients are notified and what is reviewed afterward. The path is agreed before launch.
-
Documentation and handover
Clients receive architecture diagrams, data-flow notes, system documentation and known limitations. The goal is a system the client’s team can own.
Frameworks
Frameworks we align with.
We use public frameworks to structure the work: AI risk, application security and accessibility. They inform our decisions and give clients a common language.
NIST AI Risk Management Framework
We use its functions — govern, map, measure and manage — to organize AI risk work and assign ownership.
OWASP Top 10 for LLM Applications
We design and test against the failure modes it catalogs, including prompt injection, sensitive information disclosure, excessive agency and supply-chain risk.
WCAG 2.2 Level AA
We design interfaces for WCAG 2.2 AA, with keyboard access, visible focus, sufficient contrast and reduced-motion support.
Alignment is a design commitment, not a credential. We do not display framework logos or imply endorsement.
Security
Report a vulnerability.
If you believe you have found a security issue in this website or in a system we operate, email [email protected] or [email protected]. Include enough detail for us to reproduce and verify the issue.
Please do not access, change or disclose data that is not yours. We acknowledge reports within three business days and will tell you what we can about next steps. If you are a client, please also contact your engagement lead through your usual channel.
Email the teamMachine-readable
Our security contact details are published in a security.txt file that follows RFC 9116.
Read our security.txtStart a conversation
Let’s turn AI ambition into systems you can rely on.
Tell us where you are today. We reply within one business day with a candid view of what is worth building — and what is not.
Or call +1 (714) 900-0052