Work
Problems we have solved
Anonymised patterns from our work, client names withheld. For research and open source, see Labs.
AGENT PLATFORM
Agents kept within bounds on a platform for many clients
2026 路 Technology services 路 Agent layer, designed and built
- Live in production for a global enterprise client
- 1 shared core, with an isolated deployment for each client
The problem
A company runs one AI platform for many enterprise clients. Its agents answer questions and take actions across email, chat, call notes and CRM data. One client's data must never reach another client's agents.
What we did
We designed and built the agent layer with the company's engineers. We set which actions wait for a person's approval, gave each tool its own permissions, and kept each client's data apart. Open protocols connect agents to tools and other systems.
Result
It runs in production for a global enterprise client. Each new client gets its own isolated deployment and agents on the shared core.
LAUNCH READINESS
An independent review of a consumer AI assistant before closed beta
2025 路 Consumer technology 路 Launch-readiness review
- 4 months, in a phased plan
- Go/no-go decision for the closed beta, informed by the review
The problem
A consumer AI team was about to open its assistant to a closed beta, unsure of its answer quality, security and regulatory readiness. It wanted an independent view before deciding.
What we did
We reviewed the whole system end to end. We recommended changes to how it finds and ranks information, a record of consent and licensing for its content, protection against AI-specific risks, reliability targets, and measurable bars to clear before launch.
Result
A written review and a phased four-month plan, which fed the team's go/no-go decision for the closed beta.
FEASIBILITY REVIEW
An independent feasibility review of an AI layer before a funding decision
2026 路 Healthtech 路 Independent feasibility review
Our role: Independent technical advisor, reviewing the proposed AI layer at arm's length.
- 5 components assessed, each rated low, medium or high risk
- Monitoring not diagnosis: the purpose reframed to fit the evidence
The problem
An early-stage healthtech team applied for a public research loan for a radar system that monitors elderly people at home. The funder needed independent evidence that the AI parts could be demonstrated within a proof-of-concept timeline.
What we did
We assessed only the AI layer. We reframed the purpose to what the evidence could support, monitoring and alerting rather than diagnosis, and re-scoped the proof of concept to match.
Result
A confidential feasibility report rating the five AI components low, medium or high risk, a validation plan with measurable acceptance criteria, a GDPR-aligned privacy-by-design data plan, and a phased roadmap for after funding.
A similar problem?
The first call is 30 minutes. We will tell you whether it matches what we do.
Book a 30-minute call