Operating principles
Each conviction maps to a category of failure we have watched derail enterprise AI rollouts — and to the guardrail that replaces it.
01
We build proprietary evaluation suites tuned to legal, medical, and financial risk surfaces. Generic leaderboards do not survive contact with a regulated workflow, so we measure the dimensions auditors will actually ask about.
02
Jailbreak attempts, prompt injection, and data-leakage probes are baked into every engagement before a model reaches staging.
03
Continuous regression and drift monitoring run against your actual traffic, so model and prompt changes never silently degrade user-facing behaviour.
04
Security guardrails, retention controls, and audit logging are configured before evaluation begins, not retrofitted at procurement review.
The engagement
Engagements are founder-led and stack-agnostic. We adapt to your model mix — Claude, GPT, open-source, custom — and to your existing CI, observability, and incident-response surfaces. Methodology and tooling transfer with us; you do not end up renting a dependency.
01 · Discovery
Workshop the agent's intended use, the regulated obligations around it, and the failure modes a stakeholder would refuse to accept in production.
02 · Evaluation
Compose a domain-specific eval suite covering accuracy, latency, cost, factuality, and policy adherence — graded against your own gold-standard data.
03 · Hardening
Attack the system: jailbreaks, prompt injection, exfiltration, tool misuse. Every finding routes back into the harness so the regression never returns.
04 · Production
Live drift detection, decision logging, and continuous regression testing keep the deployment trustworthy as models, prompts, and traffic evolve.
Why production-first
Most AI tooling optimises for the demo. We optimise for the on-call rotation. When an agent answers a regulated question wrong in production, the consequences land on legal, security, and the engineer who shipped it — not on the vendor that benchmarked the model.
Our methodology starts from that asymmetry. Evaluation suites are written against your own gold-standard data. Red-teaming runs before staging, not after a customer escalation. Drift detection and regression tests live in the same pipeline as the model itself, so a prompt or weights change cannot quietly regress a metric an auditor depends on.
We share our findings, our tooling, and our threat models. The goal is for your team to leave an engagement able to operate the harness without us — not to keep you on a retainer.
Compliance posture
Every engagement runs on infrastructure aligned to the certifications below, with documented retention, encryption, and access controls. Compliance is a precondition for our methodology, not an afterthought at procurement review.
Continuous controls audit covering security, availability, and confidentiality of every evaluation pipeline.
Information security management system aligned to ISO/IEC 27001, with documented risk treatment for every workload.
Lawful basis, data minimisation, retention, and DSR workflows defined before any production data is processed.
Architecture for protected health information: BAAs, encryption-at-rest, audit logs, and de-identification on the eval surface.
Founder-led engagement, scoped in a 30-minute conversation. We map the failure modes, the regulatory surface, and the evaluation gap before you commit to a pilot.