Generative AI development
Language models are good at drafting, extracting and summarising over your own documents. They are poor at anything that must be exactly right every time. We build the first kind and talk you out of the second.
Ask for a two-week pilot Read where it pays
We have no published generative AI case study yet. This page says what we would do and how we would prove it.
| Can a human check the output fast Cheap to verify | Good sign |
|---|---|
| Is a near-miss still useful Draft, not decision | Good sign |
| Volume high enough to matter Saves real hours | Good sign |
| Must it be right every time Compliance, pricing, safety | Bad sign |
| Would a query or rule do it Cheaper and explainable | Bad sign |
The honest question first
The technology is genuinely useful. The failures come from pointing it at the wrong task, not from the models being weak.
Do not use it for
We will tell you to build a query or a rule instead, and it will cost you less.
Worth building for
Every one of these is checkable by a human in seconds, which is what makes them safe.
Plain definitions
Get these right and the pilot is two weeks. Get them wrong and you are fine-tuning a model to solve a search problem.
Retrieval augmented generation means the model answers from documents fetched at the moment of the question rather than from what it absorbed during training. Your content gets indexed, the relevant passages get pulled, and the model writes an answer over them.
It matters for three reasons that all reduce to the same thing. The answer can cite its source, your content stays yours instead of being baked into a model, and updating a document updates the answer immediately.
Prompting alone is right when the task is generic and the model already knows how. RAG is right when the answer depends on your own content, which covers almost everything a business actually wants. Fine-tuning changes how the model behaves, not what it knows.
People reach for fine-tuning to solve knowledge problems and it does not work. It teaches format, tone and a narrow skill. We use it rarely, never first, and only after retrieval has been tried and measured.
What we build
Where the content already is
The task matters less than the corpus. If one of these piles is large, searched often and painful to search, there is probably a case here.
How the work runs
Days 1 to 3
The narrowest task with real volume behind it, and agree how a good answer is recognised.
Days 4 to 7
Real questions with known good answers, written with your people. Without this there is no way to tell progress from noise.
Days 8 to 12
Retrieval over your content, citations, permissions, then run the evaluation set and report the score plainly.
Day 13 to 14
Worth building, worth narrowing, or not worth doing. All three are acceptable outcomes.
Technology
Six parts. The first three are ordinary engineering, and they are where these projects fail far more often than at the model.
Who does the work
The failure mode in this work is a convincing demo. Twenty engineers with no sales layer means the person showing you the pilot is the person who has to live with it in production.
You meet whoever runs your pilot. Ask them for the evaluation score rather than a demonstration, and see whether they have one.
How we work
The models differ in one thing: how much of the management you hand over. Everything else, including who owns the code, is the same in all five.
Changing model later is normal. Augmentation into a dedicated team is the common direction, and the people stay.
Engineers join your team and work inside your process, your board and your code review.
A team that works only on your product, with a lead on our side running delivery.
A long-term engineering site under your standards, where we carry recruitment, HR, payroll and equipment.
04
One price for a scope that is genuinely settled, with the overrun carried by us.
Everything in the centre model, with the whole team moving into your own Vietnamese entity on an agreed date.
FAQs
Nothing we can point you at publicly. We would rather tell you that than dress up a demo as a track record. What we do bring is the document and permission plumbing these features sit on, which is where most of them go wrong.
Not unless you choose a provider that does that and accept it. We deploy in your tenant and configure providers so your content is not used for training.
They happen, which is why we only build features where a person can check the output quickly and a near miss is still useful. Anything that must be exactly right gets a query or a rule instead.
Rarely, and never first. Retrieval over your own content solves most of what people expect fine-tuning to solve, at a fraction of the cost.
Whichever fits your data policy and budget, including models you host yourself. We are not tied to one.
You do, including the evaluation set, which is often the most durable thing produced.
The two-week pilot is fixed price and ends in a score rather than a slide. Taking a proven pilot into production, with permissions, logging, evaluation in the release process and a real interface, is a project of its own. Anything wider, across several document sets and several teams, runs longer again. How it is billed follows the engagement model above.
This is the question nobody answers and it is a design decision, not a surprise. Cost per answer is driven by how much context you retrieve and which model you send it to, and both are adjustable. We measure it during the pilot and report it alongside the accuracy score, so you decide with the running cost in front of you rather than after the first invoice.
Narrow ones, where a person approves anything that changes a record. We would argue against an agent that acts on its own in any process with a compliance or financial consequence, because the failure mode is not a wrong answer you can see, it is a wrong action already taken. Retrieval with a human deciding covers most of what people want from agents at a fraction of the risk.
If it is checkable in seconds and happens hundreds of times a week, it is probably worth two weeks of proving. If not, we will tell you.
