Five stack reports, one method

Before we quote a build, we write down the layers the finished system needs, what a broken layer looks like from the outside, and what it costs at three sizes. The five reports are those notes, cleaned up for anyone to read. Together with the Eval Spec and the State of Eval-Driven AI, they make up the seven free reports we publish.

The five reports

The layer model

Each report names the layers a working system of that kind has, whatever its size. A hospitality stack has four, a marketplace six. Scale changes how much of a layer is shared or automated; it does not add or remove layers.

Every layer comes with one observable signal that it is broken, written so an operator can check it without opening any code: a message reaching the wrong person, a draft rewritten from scratch, a number nobody knows the age of. The diagram at the top of each report puts the layers and their signals side by side.

Each report then shows the same layers at three scales in one table, lists what we would not build at that stage, and names the live engagement behind it, or says plainly that there is none yet.

Reading the cost bands

The bands in the reports are our quoting bands, the same ones on the services pages. They are not averages from a sample of client projects. They cover design, engineering and launch, and exclude what you keep paying for afterwards: payment fees, the CRM licence, model and API spend, hosting.

• Discovery Sprint. Two weeks at a fixed price of CHF 3,500. It ends with a written scope, or a written recommendation not to build.
• One surface, such as a venue website: CHF 8,000 to 12,000. With booking, a light CRM and automation behind it: CHF 12,000 to 18,000.
• A platform, such as a marketplace or an agent system across a team: CHF 15,000 to 25,000, with a ceiling of CHF 45,000 when the scope grows.
• Timing. A focused MVP can be ready in two weeks. A scoped product can launch in around two months. We agree the timeline after reviewing your requirements.
• After launch, a monthly retainer from CHF 800 for iteration, tuning and fixture refresh.

Builds are priced per project, never by the hour. The total is fixed in the proposal and paid in milestones, and a smaller budget gets a smaller scope, not a lower rate.

Where a project lands inside a band depends mostly on integrations: a till or booking system with an awkward API, a legacy CRM, several languages, an audit trail. When a scale in a report says “scoped per case”, we have not built it yet and will not guess a number.

The self-audit

Every stack report ends with twelve yes-or-no questions. Answer them from what your tools do today, not from what is planned, and ideally with the person who does the work in the room. The checklist on each page keeps a running score; the bands mean the following.

• 0 to 3: People are the integration layer. Most of the work between the tools is done by hand. Start with the lowest-numbered question you answered no to; the early questions are the foundations.
• 4 to 7: Layers exist, seams show. The parts are there but they do not talk to each other yet. Look at which layer your no answers cluster in. That layer is where the week leaks time.
• 8 to 10: A working stack with gaps. The stack carries the week. Close the remaining gaps before adding anything new.
• 11 to 12: The stack does its job. Run this again after your busiest season. Scores drop quietly when nobody is looking.

A low score is a map, not a verdict. The questions are ordered so the early ones are the foundations the later ones depend on.

The rules the reports assume

Our live engagements are Swiss businesses, so the reports assume Swiss and EU rules. These notes apply to all five.

• Data protection. The revised Swiss FADP and the EU GDPR both apply to guest, customer and document data. Build consent and retention rules into the memory layer from the first version. For regulated clients, storage and model inference stay in Swiss or EU regions.
• Languages. The German-speaking cantons work in German and English, Geneva and Vaud in French and English, Ticino in Italian. Labels, drafts and extraction prompts are kept per language rather than as one multilingual version.
• Payments. Swiss consumer checkouts usually need TWINT next to cards. For B2B, invoicing and bank transfer are often enough.
• Calendars. Public holidays differ by canton. Operations and matching layers should follow the canton, not one national calendar.
• EU AI Act. Obligations phase in through 2026 and 2027, and some uses may be classed as high-risk. The artefacts an auditor asks for (an eval harness, traces, a rubric per workflow) are cheaper to build from day one than to retrofit. The method is in the Eval Spec.

The work behind the reports

Two reports rest on builds you can look at. Incontro Bar, an aperitivo venue, is live: reservation flow, events, email automation and a lightweight CRM and CMS. Dreilokale, a Swiss event-location matching platform, is live: matching engine, CRM, CMS and email automation, with no AI workflow in production.

The CRM intelligence, document intelligence and dashboard reports describe the shape we build to. None of them has a named live engagement yet, and each says so. The work is done by three people: Edward Salvatierra (vision, AI systems, engineering), Alexander Koch (strategy, scope, timeline) and Kevin Silva (design and visual systems).

Versioning

Each report carries a version number, a first-published date and a last-reviewed date. A new layer, a new scale or a change to a cost band bumps the major or minor version; wording fixes do not. All five were cut down on 10 October 2026, when the shared method moved to this page, and last reviewed on 12 October 2026, when the cost bands moved to our CHF price list and the build times to qualified timing: a focused MVP can be ready in two weeks, a scoped product can launch in around two months, and the timeline is agreed after reviewing requirements.