AI product development, engineered for production
Quantum Craft does AI product development for funded startups and UK companies: MVPs, SaaS platforms, mobile apps and AI features inside existing products. AI does much of the coding, and a senior engineer reviews every change against a written specification. A single-workflow product built to production standard typically costs £35,000 to £80,000 over 6 to 12 weeks; larger platforms such as BlueWave and Karu each had a core build of about 14 weeks.
AI-accelerated product engineering is building software where AI does the wide, repetitive work and senior engineers own the specification, the architecture and every merge.
- Specification sprint: £4,500 for one week, £8,500 for two, excluding VAT
- You keep the specification
Last reviewed
- Focused MVP
- 6 to 12 weeks
- MVP band
- £35,000 to £80,000
- Platform core build
- About 14 weeks
- Every change
- Senior review and automated tests
Apps, SaaS platforms and MVPs we build
Most of our work takes one of four shapes. Where a claim can be checked, it links to the platform behind it.
MVPs and SaaS platforms
A first version with one workflow done properly, or a multi-tenant platform with roles, billing and an admin console behind it, like BlueWave. We keep customers apart in the data layer, where a mistake fails loudly, instead of trusting the interface to hide one account from another. On Karu, every query filters by practice before any other condition.
Mobile apps
Native iOS and Android from one Flutter codebase, as in NForge's coach app, and offline-first field apps that keep working in a basement plant room, then sync when the engineer gets signal back. We have written up the offline architecture we use.
AI features inside products
Assistants that answer from your own records, document extraction, agents that act through a small set of tools, and generation pipelines whose output is checked before anyone sees it. A B2B logistics marketplace we built runs five LLM surfaces in production. Before you commission an agent, check the job needs one: a plain workflow is often cheaper to run and easier to test.
Taking over existing code
Code another team wrote, including prototypes generated with AI app builders such as Lovable, Bolt.new or Replit. We read it before we quote, keep what is sound and replace what is not. For apps built mostly by prompting, vibe code rescue starts with an AI-built app audit at a fixed £3,500, excluding VAT.
Where AI speeds up software development, and where it doesn't
AI saves us the most time on wide work that follows a pattern. It saves almost none on the decisions that set the cost of everything after them, and that is where an engineer's hours go.
| Kind of work | Time AI saves | What still takes an engineer's time |
|---|---|---|
| Wide mechanical changes | A lot. A new field carried through every layer, or a concept renamed across hundreds of files, is mostly typing | Reading the whole change against the specification before it merges |
| Tests from acceptance criteria | Most of the drafting | Checking that each test fails without the code it covers |
| Screens and endpoints on an existing pattern | Most of the first draft | The edge cases the pattern does not cover |
| Documentation | Some. First drafts of API references, runbooks and migration notes | Correcting the draft and cutting what a reader will not need |
| Tenancy, sign-in and payments | Little | The design, and the slowest review anything gets |
| Architecture and the data model | Almost none | The decision, written down with its reasons |
| Subtle bugs in money, concurrency or sync | Almost none | The diagnosis, and usually the fix |
| Anything that turns on intent | None | Asking the person who knows what the business meant |
The saving is real, and it lands in the typing. A wrong assumption costs the same whether a person or a model typed it. How we work lists the checks every change passes before it merges.
How much does an MVP cost in the UK, and how long does it take?
Every build starts with a specification sprint: the written specification, the architecture, an estimate and the risks. It costs £4,500 for one week on a single workflow, or £8,500 for two weeks on a platform, excluding VAT. The document is yours. After that, these are the shapes we have built and what a build of each shape costs.
| Shape | Our example and its scope | Time | Indicative band |
|---|---|---|---|
| Single-workflow MVP | One workflow built to production standard. The bottom of the band has one user type and no integrations; the top adds payments, an admin surface and a second audience | 6 to 12 weeks | £35,000 to £80,000 |
| Practice platform | Karu: scheduling, records, notes and video | About 14 weeks of core build | £90,000 to £140,000 for this phase |
| Two-sided marketplace | A B2B logistics marketplace: the first transactable phase | Phase one of a multi-year programme | £125,000 to £170,000 |
| Field-service compliance platform | BlueWave: web app, offline mobile app, scheduling and billing | About 14 weeks of core build to first release | £150,000 to £250,000 to first release |
| Platform with native apps and payments | NForge: web and native mobile apps, Stripe Connect payments and live video | 30 to 40 weeks | £260,000 to £340,000 |
| Data platform programme | TIYO: a crawl engine, an event store, analytics and several products | Two to three years | £350,000 to £550,000 across the programme |
All bands exclude VAT and costs paid to third parties such as hosting and infrastructure, at London rates as of mid-2026. They are bands for the shape, not an invoice: your figure comes from the specification. Once the product is live, plan for 20 to 30 per cent of the build cost a year to run it and keep changing it.
Below the MVP band you are buying a prototype. That can be the right purchase for testing an idea, provided nobody later mistakes it for the product.
Adding AI features and AI agents to your product
A model that impresses in a demo tends to fail in production in familiar ways. It states things that are not true, and it costs more each month than anyone planned. Then the provider retires the version it was tested on. We build the controls in from the first release. Each one below comes from a product we built.
Evaluation before launch
An AI feature gets a fixed set of test scenarios, run again whenever the prompt or the model changes. NForge's programme generator has an evaluation harness that runs the real pipeline over synthetic and replayed intakes, and plants known defects to check that its verifier catches them.
Prompts instruct, code enforces
Anything that must be true is checked by code on the server. NForge computes calories from macros instead of trusting the model, and around thirty validation gates run before a coach sees a programme; one that cannot be repaired goes to a person. On a B2B logistics marketplace we built, the assistant cannot present a warehouse unless a real tool call returned it.
Agents act through narrow tools
An agent gets named tools with a declared effect, rather than open access to a database. The same marketplace exposes its operations to AI agents through a 50-tool MCP server, and every tool is classified as read, create, mutate or destructive.
Cost you can see per call
We log what every AI call costs, by call and by model. NForge's public trial has rate caps, a daily budget breaker and a kill switch, because a free demo running on a paid model is an open tab.
More than one model
The main model is chosen in a bake-off on results and cost, with a fallback ready. The logistics marketplace runs several providers behind weighted A/B tests. When a price rises or a model is retired, we re-run the evaluation instead of rewriting the feature.
Personal data stays where it should
For AI over sensitive personal data we start from Karu's rules: inference in the UK or EU, terms that bar training on your customers' data, a data protection impact assessment first, and features that ship switched off. Karu's clinical drafting layer is designed to those rules and has not shipped yet.
If the product is used in the EU, whether you are the provider or the deployer decides your duties under the EU AI Act.
Will AI-written code pass technical due diligence?
It will if the evidence is in the repository. Investors now ask how AI-written code was checked, and "we were careful" is not an answer anyone can verify. This is what we give them to check.
Tests that fail without the change
We write each test to fail if the code it covers breaks, so a green build means something. BlueWave carries more than 1,000 automated tests, and NForge has 1,258 backend tests.
Decisions with their reasons
Architecture decisions go on record with their reasoning, so a reviewer can see why the system is shaped the way it is. BlueWave has 10 architecture decision records.
Senior review before every merge
A senior engineer reviews every change before it is merged, including the parts AI drafted, and checks it against the specification.
Code that is yours
Ownership passes to you on full payment, and the repository can sit in your own organisation from the first commit, so a reviewer sees the whole history. Documentation is written alongside the code, so nothing has to be pieced together for the review.
Reviewers work from a known list, and the technical due diligence checklist investors work from shows what they open first.
One number from each build
Five platforms designed and built end to end: three live, two in build. Each figure below is published in the platform's own case study.
- 1,000+automated tests
- 1,258backend tests
- 0client logins to steal, because clients never get one
- 5LLM surfaces in production, with an evaluation rig
- 22retailers tracked across 13 countries
Each case study puts its number in context, with the scope of the build and its price band.
What the platforms we build run on
These are the technologies under the platforms we have built recently. We choose per product, and we would rather use a dull tool we know well than learn a fashionable one on your budget.
| Back end | .NET 8 and FastEndpoints, with CQRS and event sourcing where the history matters |
|---|---|
| Data | PostgreSQL, Elasticsearch, Redis and Bigtable |
| Web | Next.js 16, React 19, TypeScript and Tailwind |
| Mobile | Flutter for native iOS and Android, and offline-first web apps on IndexedDB |
| Infrastructure | Google Cloud: Kubernetes in the London region, Pub/Sub and Cloud Storage |
| Identity, payments, video | FusionAuth, Stripe and Stripe Connect with Bacs Direct Debit, and LiveKit |
| AI in products | Model APIs from more than one provider, evaluation rigs and MCP servers |
Payments get the most care. We have written up four Stripe Connect billing decisions that are hard to reverse.
Who a build with us suits, and who it doesn't
Best for
- Funded founders who need a first release that survives real users and technical due diligence
- UK companies adding AI features, such as assistants, search or document extraction, to a product they already run
- Products that take payments or hold regulated data, where the hard parts need senior hands
- Teams that want the specification, the code and the accounts in their own name from the first commit
Not for
- Testing an idea cheaply. A prototyping tool may be the better purchase for that
- Deciding what to build. If that is still open, AI consultancy is the better first step
- Building without a written specification. We write one first, and it is not an optional step
- A large team from the first week: we are a small senior team
Not sure what to build yet?
Then a build is the wrong place to start.
If the open question is what to build, or whether AI belongs in it at all, start with AI consultancy. The advice comes from the engineers who would do the building, and sometimes it is to buy an existing product, or to wait. That is cheaper to hear before you spend money on code.
Questions buyers ask before a build
How much does MVP development cost in the UK?
A focused MVP, one workflow built to production standard, typically costs £35,000 to £80,000 excluding VAT, over 6 to 12 weeks. The bottom of the band has one user type and no integrations; the top adds payments, an admin surface and a second audience. The specification sprint that comes first costs £4,500 or £8,500.
How long does it take to build an AI MVP?
A focused MVP takes 6 to 12 weeks, after a specification sprint of one to two weeks. BlueWave and Karu, both bigger than an MVP, each had a core build of about 14 weeks. An AI feature adds time for evaluation, which we plan from the first week instead of squeezing it in before launch.
Can you add AI features to our existing product?
Often that is a better first step than a new product. We read your codebase first, then build the feature inside your existing permissions and data rules, with an evaluation set in place before launch. Document extraction and assistants over your own records are the usual starting points.
How do you keep our data secure when AI is involved?
While we build, AI works on your code only through business accounts whose terms bar training on it. It never holds production credentials, and client personal data stays out of it. Inside your product, AI features use providers whose terms bar training on your data, with processing in the UK or EU where the data needs it.
If AI writes the code, why isn't it cheaper?
Because typing was never most of the cost. AI saves real time on wide, repetitive changes. Specification, architecture, review and testing take as long as they always did, and they are what keep a product working after launch. We spend the saved hours there, which is why our bands look like a senior team's.
Who owns the code you build with AI?
You do, including the parts AI drafted. The code we write for your project is yours on full payment, and it can sit in your own repository from the first commit. Our own tools and methods stay ours. UK copyright law on wholly AI-generated work is under review, which is one more reason to put ownership in writing.
What happens after launch?
Plan for 20 to 30 per cent of the build cost a year to run the product and keep changing it. Ongoing work usually runs as a retainer; BlueWave's is two to three days a week of senior time. If you would rather take the product in-house, the handover is written down and the code is documented for your team.
Which AI model will our product use?
Whichever wins an evaluation on your own data, weighing quality, cost and where the data is allowed to go. We keep a second model ready, so a price change or a retirement notice from one provider does not strand the feature. You see the evaluation results before we commit to a model.
Tell us what you're building.
Tell us about your product. On a 30-minute call we'll ask who it is for and what it has to do by when, then tell you what a first release would take.
What happens after you get in touch
We reply within one working day
By email, to arrange a time for the call.
A free 30-minute call
With a senior engineer, not a salesperson, about what you are building or deciding.
The fee in writing first
If a first step is worth taking, you get its exact fee in writing before any work starts.