Startups ~6 weeks Fixed scope Your repositories

Your AI MVP in six weeks.

Not a prototype that demos well. A working slice of product, running on real data, that real users can try and you can learn from.

THE SHAPE OF THE WORK

Week 1

Hypothesis and architecture

What single question does this MVP answer, and what would count as a yes? Everything else gets cut. This week produces a written hypothesis, an architecture, and the evaluation approach.
- —One-sentence hypothesis with a success threshold
- —Architecture and model selection
- —Data access established — usually the long pole
Weeks 2–4

Build the core loop

The one thing the product must do, done properly: retrieval or agent loop, the data pipeline behind it, and enough interface for a real user to use it unaided.
- —Core AI system on real data
- —Minimal but real product surface
- —Evaluation harness running from the first week of build
Week 5

Sharpen against the eval set

This is where the quality actually arrives. Retrieval tuning, prompt iteration and failure-case handling, measured rather than eyeballed.
- —Measured iteration against the eval set
- —Cost-per-interaction modelled at projected volume
- —The failure cases handled explicitly
Week 6

Ship it to users

Deployed, instrumented and handed over with documentation. You end the engagement able to run and change it without us.
- —Deployed with logging and usage analytics
- —Handover documentation and architecture notes
- —A written view on what to build next — and what not to

WHAT KEEPS IT TO SIX WEEKS

One hypothesis, ruthlessly

Every MVP that overruns does so because it grew a second goal. We write the hypothesis down in week one specifically so that later requests can be measured against it and deferred rather than absorbed.

Data access starts on day one

The most common cause of a slipped AI MVP is not model work — it is waiting three weeks for API credentials. We start that clock before anything else, and if it is going to be slow we design around it.

Boring infrastructure

Managed services, a conventional stack, no novel infrastructure. An MVP is not the place to also learn a new deployment platform. The interesting risk should be in the product, not the plumbing.

Evaluation from the start

An eval set built in week one turns week five from anxious prompt-fiddling into measurable iteration. It is the highest-leverage hour in the whole engagement.

QUESTIONS

QUESTIONS — 5

Typically €30–60k for a six-week engagement, driven mostly by how many systems we have to integrate with and how much work the data needs before it is usable. That is a rough estimate from past work rather than a fixed price; we quote each project on its own scope after a scoping call.

Then the idea needs splitting, and that is worth knowing before you start. Almost every ambitious product has a six-week slice that tests its riskiest assumption. If yours genuinely does not — which happens with heavy regulatory or hardware dependencies — we will tell you and propose a different shape rather than pretending the timeline works.

No, but it helps enormously if someone on your side can make product decisions quickly. The main thing that slows an MVP is not engineering, it is waiting for a decision about what the product should do. If you have engineers, we pair with them and hand over knowledge as we go.

It is real software running on real data with real users, and it is deployed properly. It is not hardened for scale, because hardening for a load you have not proven you will get is how startups waste money. What we do commit to is that it is not designed as a dead end — an architecture intended to be extended rather than thrown away.

Whatever suits you. Some clients take it in-house immediately, which is a good outcome and one we prepare for. Some continue with a fractional arrangement. Some stop, because the hypothesis was disproved — which is the MVP working exactly as intended, and considerably cheaper than finding out in a year.

Keep reading

Prototype to Production

→

Prototyping

Web Application Prototypes

→

30-minute call · no deck