I'll tell you when it isn't AI
A surprising share of "AI problems" are a badly designed form, a missing integration or an unclear approval rule. If a week of process fixing beats six weeks of model work, that's the recommendation you'll get.
Home/Approach
How I work
Four stages with a genuine exit after each one. You keep everything produced up to the point you stop — which is the only structure that keeps a consultancy honest about whether the next stage is worth buying.
Discovery sprint
I watch the work happen. That means sitting with the people doing it, not only interviewing the managers describing it — the gap between those two accounts is usually where the opportunity is.
- End-to-end process mapping with time and error hotspots
- Data readiness assessment per candidate use case
- Scored shortlist: value, feasibility, risk, dependencies
- Costed roadmap and a written recommendation
Exit here and you keep: the process maps, the scored shortlist and the roadmap — useful whether or not I build anything.
Pilot build
One use case, built against real data, with the definition of "working" agreed before I start. I write the test set first so success is a number rather than an opinion in a steering meeting.
- Evaluation set defined and signed off up front
- Working system on real data, used by a small group
- Weekly demos — you see progress, not a reveal at the end
- An honest read on whether it beats the status quo
Exit here and you keep: the code, prompts, evaluation set and documentation, in your repositories.
Production rollout
The distance between a working pilot and a system a whole department depends on is mostly hardening, integration and change management. I roll out to one team, fix what surfaces, then widen.
- Integration into the tools people already use
- Access control, audit logging, failure handling
- Monitoring for output quality, cost and latency
- Training, comms and the adoption work that decides success
Exit here and you keep: a production system running in your own cloud accounts, fully documented.
Handover or run
Two honest options, and I'm not incentivised toward either — the retainer is priced so that it genuinely doesn't matter to me which you choose.
- Handover: hands-on training, runbooks, governance policy, and a support window while your team takes the wheel
- Managed: I watch quality, cost and drift, handle model upgrades, and send a monthly report
Principles
What I hold to, even when it costs me the work
Evaluation before enthusiasm
Nothing ships on the strength of a good demo. I define the test set before building and report against it, including the cases where the system underperforms.
Your infrastructure, your IP
Everything runs in your accounts and lands in your repositories. There's no platform of mine to license and no switching cost engineered into the relationship.
One person, small scope
The person who scopes your project is the person who builds it — there is only one of me, which is the point. I'd rather run a couple of projects properly than staff a bench.
Start with two weeks, not two years.
A discovery sprint is a small, fixed commitment that ends with a decision you can defend either way.