How we work

From your first call to production, in four stages.

Diagnostic, Discovery, MVP, Production. Fixed prices where the scope is fixed; bands where it is a judgement call. Discovery fee credited.

Last updated:

The engagement

The four stages

  1. 01

    Diagnostic call

    30 minutesfree

    A half-hour call with the person who would actually run your project. No sales handoff, no junior on the other end. You describe the problem; we ask questions until we understand the operational reality, the data that will feed the system, and the team that will use the result. By the end, you know whether AI is the right tool for this job. We send a written summary either way.

    Best forany conversation worth having before paying anyone.

  2. 02

    Discovery / proof of concept

    1-2 weeks€2,000-€3,000 fixed

    A written roadmap plus a small working demo on your data. The roadmap covers scope, recommended architecture, integration points, risks, and a real cost-to-build figure for the MVP. The demo is intentionally small - usually one user-facing scenario, end-to-end, built on a snapshot of your real data so you can poke at it.

    Deliverables
    roadmap PDFworking demodata-handling memoMVP cost estimate.

    Best forgetting to a confident yes or no on whether to build, with enough information to decide.

  3. 03

    MVP build

    4-8 weeks€5,000-€15,000

    The first production-ready system. Not a prototype, not slideware - a system handling real work for a small group of pilot users, with monitoring, error handling, audit logging, and a runbook. We build with weekly checkpoints; you can stop at the end of any week with a usable handover.

    The Discovery fee is credited against your first MVP invoice. If you spent €3,000 on Discovery, the MVP is €5-15k minus €3,000.

    Deliverables
    deployed systemmonitoring dashboardrunbookweekly demo recordingshandover document.
  4. 04

    Production + maintenance

    ongoingquoted per project

    A live system with us on call. Monitoring, iteration on what we learn from real usage, periodic security review, and a clean handover path the day you want to bring the work in-house. We bill monthly on a flat retainer or a quoted scope, depending on the system; both are spelled out in writing before either side signs.

    What we will not do: lock you in via a stack you cannot hire for. The system runs on mainstream technology your team can take ownership of; the runbook covers exactly how.

Two delivery models

Cloud or Local

Every project ships in one of two flavours. The choice is yours; we recommend a fit during Discovery.

Cloud API
  • Models: Claude · GPT · Gemini · Llama via the providers' APIs.
  • Pricing: per-token variable. Pass-through at cost plus a fixed margin, or a monthly cap.
  • Iteration: fastest access to new model releases.
  • Data flow: sent to the provider for inference under their DPA.
  • Best for: prototyping, variable workloads, frontier reasoning.
Local LLM, hosted
  • Hardware: NVIDIA DGX Spark in our Sofia office.
  • Models: Qwen 3 FP8 · Gemma 4 · custom fine-tunes.
  • Pricing: fixed monthly fee. No per-token surprises.
  • Compliance: GDPR · EU residency · NIS2-ready · zero training-data leakage.
  • Best for: regulated industries, sensitive data, high-volume workloads, defensible data-residency story.
FAQ

Common questions

Why is Discovery paid?
Because if it is free, we cannot afford to do it well. A €2-3k fixed Discovery means we put our best person on it for the time it actually takes, write down what we learn, and credit the fee against your MVP invoice if you choose to build.
What if AI is not the right fit?
We say so in writing - sometimes before invoicing the Discovery, sometimes during. A meaningful share of our diagnostic calls end with "do this without AI; here is how." That is the call we would want someone to make for us.
Can we own the code?
Yes. Default contract terms transfer ownership of the application code we write to you on payment. Models we fine-tune on your data are yours. Frameworks we build on belong to their authors (open-source, mostly).
How does the local LLM compare to the API on cost?
The crossover is usually around 8-12M tokens per month. Below that, the API wins. Above, the local LLM wins. We model it for you with your usage estimates during Discovery.
Where are you based and who do you work with?
Sofia office; our clients are global. Most of the work is remote, with on-site visits where it adds value. We work with companies in the US, in Europe, and elsewhere.
What happens if you and we disagree on scope mid-project?
We re-scope in writing before any work continues. New scope, new estimate, your decision to greenlight. We do not bury changes in a monthly invoice and we do not abandon a stage halfway through to chase the next one.
Does our team get trained on the system you build?
Yes - knowledge transfer is part of every engagement. Reading the code we write, debugging it, taking ownership at handover. Standalone team-wide AI training is a separate offering on the Services page.
AI that already runs

Start with the free 30-minute diagnostic.

No sales pitch - a working call where we figure out together whether this is worth pursuing. Written summary either way.

Book a callRead a case study first
Live · powered by Gemma-4 · running on our hardware in Sofia