Services
Most AI budgets are spent before anyone has worked out which project deserves the money.
What follows is what we actually do, grouped by what you are trying to get done. All of it runs through the same three steps, and the list of what we turn down is on this page too.
Work delivered for
- Singapore Government project
01 · Identify
Work out what is worth building
Everything starts here, and for a good number of companies it also ends here, because the useful output is a decision rather than a build.
AI opportunity assessment
Two to three weeks inside your company, ending in a ranked list of what is worth building and a longer list of what is not.
We sit with the people doing the work rather than the people describing it, read how a task actually moves, and price the friction in hours and money. Every surviving candidate gets a baseline number and a cost model you can argue with.
Mostly bought by software firmsWhat you end up holding
A 12 to 20 page opportunity note, written to be read by your CFO as well as your engineers. Fixed fee, fixed end date, nothing owed afterwards.
Technical due diligence
A technical read on a company you are about to back, in 48 hours, before the committee meets.
Whether the codebase matches the story, where the team carries risk it has not priced, and what the first six months of engineering will cost. The memo states its own gaps on the first page rather than implying coverage it does not have.
Mostly bought by vcs and acceleratorsWhat you end up holding
A three to six page memo plus the questions worth asking the founder directly. The first one is free.
Fractional CTO
A technical partner on retainer for a founder who has raised money and has no CTO. Ten to twenty hours a month, starting with a written assessment in the first thirty days.
The job is to stop expensive irreversible decisions being made by accident. Choosing the stack, sizing and interviewing the first engineering hires, reading the code an agency delivered, pricing a vendor properly, and answering the technical questions your next investor will ask. It starts with a written technical assessment inside thirty days and a ninety-day plan with milestones somebody can check, because a retainer with no plan becomes a standing call that nobody can evaluate.
Mostly bought by startupsWhat you end up holding
The assessment in month one, a ninety-day plan, and a named person who reads the code and sits in the interviews. One month notice either way, no annual lock-in.
Fractional AI advisory
A technical partner for a startup that needs AI judgement without hiring an AI team to get it.
Standing input on what to build, what to buy, which model to use and when a plan is about to cost more than it returns. It works as a recurring session plus availability between them, and it suits teams whose engineers are capable and short of time to work out what is worth their week.
Mostly bought by startupsWhat you end up holding
Decisions made faster and a written record of why each one was made, so the next person does not relitigate it.
02 · Build
Build it
Software your team uses on a Monday, in your repository, under your review process. Not a prototype that dies the week after the demo.
AI features in your product
The AI part of your SaaS, built into the product your customers already pay for.
Agents, assistants, retrieval, summarisation, classification, whatever the assessment said was worth building. It ships with an evaluation set of real inputs, the failure cases handled, and a cost model that survives your actual traffic rather than a demo.
Mostly bought by startupsWhat you end up holding
The feature running in your environment, an evaluation set your engineers can rerun, and a runbook written for whoever inherits it.
Software and platform builds
The engineering around the AI, and sometimes instead of it.
Integrations, data pipelines, internal tools, the reconciliation layer nobody wants to own. On the DHL programme the largest single win was deduplicating traffic between two systems with no model involved at all, and that is a common shape. Where the honest answer is deterministic code, we build deterministic code and say so.
Mostly bought by software firmsWhat you end up holding
Working software in your stack, with the integration points and failure behaviour documented where your team documents everything else.
AI transformation
Rebuilding how work actually gets done inside the company, not just adding a feature to the product.
This is the multi-team version: mapping the processes, finding the handful worth automating, building into the systems already running, and training the people who inherit them. It is slower than a single feature and it is where the operational savings usually are. It is also the engagement most likely to conclude that your biggest win is a data quality problem rather than an AI one.
Mostly bought by software firmsWhat you end up holding
A process map with numbers on it, the builds that came out of it, and a before-and-after note per process.
03 · Prove
Make it pay
The half that decides whether any of the above was worth doing. A build nobody adopts and a number nobody recorded both return zero, so training and instrumentation sit here alongside the growth work. Two of these are product and growth engagements rather than AI ones, which is stated plainly below.
Onboarding and activation
Getting more of the people who sign up to the point where the product is worth paying for.
We read where new users stall, in the product analytics and in the support queue, and fix the specific steps losing them. Some of those fixes are AI, most are not. A guided setup, a removed field, a better empty state and a well-placed prompt all count, and we build whichever the evidence points at.
Mostly bought by startupsWhat you end up holding
The activation rate measured before and after, against a definition you agree first, and the changes that moved it.
Retention and churn
Finding out why people leave, then building the thing that keeps them.
Churn is usually several problems wearing one number. We separate them: the customers who never activated, the ones who outgrew you, the ones a competitor took, and the ones who hit a specific failure and gave up. Only some of those are addressable, and naming which is most of the work.
Mostly bought by software firmsWhat you end up holding
A churn breakdown with the addressable share separated out, and the builds that follow from it, measured against the cohort baseline.
AI training for your team
Onsite or remote. Your people leave able to do the work themselves, tested on your own tasks rather than on exercises.
Half-day sessions through to a four-week course, built around what your team actually does. Engineers get evaluation sets, cost control and the failure modes that bite in production. Everyone else gets the same short list of tasks worth handing to a model and the longer list that is not, using your documents and your tools rather than a generic example. We run it in your office or over video, and the material stays with you.
No Slides Day, the one-day versionMostly bought by software firmsWhat you end up holding
A recording, the exercises and the prompts your team wrote during the session. We test the same tasks before and after, so you get a number rather than a satisfaction score.
Measurement and instrumentation
Making a process measurable so that anything you build on top of it can be proved.
Frequently the first piece of real work. Where a process has never been instrumented there is no baseline, which means no engagement on it can ever be proved. It is unglamorous, it is usually cheap, and without it the third step of the method has nothing to read.
Mostly bought by software firmsWhat you end up holding
The number your team argued about, now recorded the same way twice, with the definition written down so it cannot drift.
Boundaries
What we do not do
A practice with no boundaries is an agency that will do anything for anyone. These are the jobs we turn down, and saying so here saves both of us a call.
- Staff augmentation. We are not a way to add three engineers to your team for six months.
- Model training from scratch. Almost nobody needs it, and the ones who do have a research team already.
- Regulated builds in healthcare or core banking where the compliance work around the system is the majority of the work. We do not carry that specialism in-house and would be learning it on your budget.
- Rescuing a project three weeks before a launch date that was never realistic. We will look at it and tell you honestly, but the answer is usually to move the date.
- "AI strategy" as a standalone deliverable. A strategy nobody builds is a slide deck, and there are cheaper places to buy one.
- Anything where you already know what you want built and simply need hands. That is a good problem and we are an expensive way to solve it.
- Fractional CTO work where the real need is a full-time CTO. Past roughly ten engineers, or where the technical decisions are daily rather than weekly, ten to twenty hours a month is a way of postponing a hire rather than an alternative to one, and we will say so.
- Off-the-shelf AI training. We only run a session built on your tasks and your documents, because a generic prompt-engineering course is worth roughly what it costs, and we do not issue certificates.
However it starts
It runs through the same three steps.
FAQ
Questions about the list
Next step
If two or three of these look relevant and you are not sure which comes first, that is the call. Working out the order is most of the value.