AI that ships, not slides.
The model is not the strategy.
Most AI projects fail before a line of code is written. They start with a demo looking for a problem, not a decision that needs to improve.
We start with the work: what decision matters, what data is allowed, who is accountable when the system is wrong. Then we prototype. Sometimes the honest answer is that AI is not the right tool. We say so.
The result is not a chatbot bolted onto a website. It is a small, evaluated system that does one job well enough to trust — and fails in ways you can see.
Small systems. Real constraints.
Six ways we help, from a first honest audit to a production system with guardrails.
Readiness audit
A structured look at your data, decisions, ownership and risk before anyone buys a GPU.
DiagnosticOpportunity mapping
Find the few workflows where AI could plausibly help, rank them by value and difficulty, and kill the rest.
WorkshopFocused prototype
A working proof in weeks, not quarters, built to answer a specific question with real examples.
PrototypeRetrieval assistant
An internal assistant that searches your documents, cites its sources and says when it does not know.
BuildWorkflow automation
Remove repetitive reading, sorting and drafting without removing the person who is accountable.
BuildTeam enablement
Workshops and working sessions that leave your team able to judge, test and maintain the system.
TrainingSix questions. Two minutes. No email required.
A quick read on whether your team is ready for a useful pilot — or needs to fix the foundations first.
Nothing you choose leaves your browser.
What we won’t do.
AI is useful and overhyped in roughly equal measure. These are the lines we keep.
No surveillance dressed as convenience.
We do not build systems whose main purpose is to watch people who did not choose to be watched.
No automated decisions without an appeal.
If an output affects a person’s rights, money or access, a human must be able to review it.
No data without permission.
If the rights are unclear, we stop. “It was on the internet” is not a data policy.
No AI as decoration.
If the simpler solution is better, we will recommend the simpler solution — even if it is less exciting.
Sefr shows its sources.
Sefr is our internal study in a narrower kind of AI: retrieval first, citations visible, uncertainty stated, and a human review step before anything leaves the workspace.
- Status
- Prototype
- Focus
- Research workflows
- Guardrail
- Citations required
- Built with
- Retrieval + evaluation
Start with the problem, not the model.
Bring us a workflow that is slow, expensive or error-prone. We’ll tell you whether AI belongs in it — and what a small, honest pilot would look like.
Book a readiness call