01
EU data residency by default
Product runtimes on our own infrastructure across the EU (Germany, France) and North America (Canada). See our security posture.
AI integration services, added to software you already run.
We add retrieval-augmented search, assistants and extraction to a working SaaS, with evaluation and cost limits so the feature stays predictable. EU-first data residency by default. A fixed first milestone, scoped and priced before kickoff. The person you email writes the code.
Updated:
Capabilities
What we add.
Four capabilities that ship inside your product, not beside it.
01
Retrieval-augmented search over your own corpus
Search grounded in the documents, rows and records your product already holds. Content is chunked, embedded and indexed in your own PostgreSQL, with pgvector or whichever vector store you already run, then retrieved at query time and cited back to the source it came from.
02
Assistants with tool use and a human-review path
An assistant that can call your own functions: look up an order, run a report, draft a reply. Every action is logged, input and output carry guardrails, and anything that touches money, customer data or a live system waits for a human before it ships.
03
Extraction from documents and uploads
Structured fields pulled out of PDFs, invoices, contracts and uploads: dates, totals, line items, parties, terms. The output is a typed record you can query and validate, not free text, and every extracted field keeps a pointer to the page it came from.
04
Evaluation and cost limits so the feature stays predictable
A fixed set of real examples from your data, run against every prompt, model and retrieval change before release. Per-request cost tracking with a ceiling, so a popular feature cannot quietly become an unbounded line on the invoice.
AI integration services are built on boring, well-understood technology with no vendor lock-in, so the stack stays yours: your repository, your database, your deployment. Nothing is proprietary and nothing is rented that you cannot take with you. Every engagement starts with a written scope and a fixed first milestone, scoped and priced before kickoff.
Starting
What we need from you.
Three things, and nothing else.
01
Access to the corpus
Read access to the documents, rows and records the feature will read. A staging copy is fine. We do not need production write access to start, and we do not need a copy of your production database.
02
An evaluation set
A handful of real examples with the answer a good one would give. Small is fine, and small and real beats large and synthetic. If you have none, we write one together in the first week and you approve it before anything ships.
03
A place to deploy
An environment we can deploy to, wired to your CI/CD, or ours if you have none. That is the whole list: no roadmap, no org chart, no long discovery phase before a written scope exists.
If the question behind your email is how to add AI to your SaaS without rewriting it, that is this service. The corpus, the evaluation set and a place to deploy are yours to give; the retrieval design, the guardrails, the cost ceiling and the EU-first deployment are ours to do.
The method is not ours alone: the measure-and-manage half of the NIST AI Risk Management Framework is where evaluation sets and cost ceilings come from, and the risk list an LLM feature inherits is the OWASP Top 10 for LLM applications.
Limits
What we do not take on.
Three requests we decline, and the reason each one fails.
01
A rewrite of a working product
Your product runs, has users, and has a roadmap. We are not the team that rebuilds it. We add one capability inside it, keep the blast radius small, and leave the rest of the codebase alone.
02
A feature with no measurable success metric
If nobody can say which number will say the AI feature worked, there is no scope to write, no evaluation set to build and no way to accept or reject the result. Name the metric before kickoff, or we decline.
03
A black-box model call with no evaluation and no cost ceiling
A prompt stapled to an API endpoint, with no test set, no traceability and no spend limit, is not a deliverable. We will not ship an AI feature that cannot be measured or bounded.
Other ways to work with the studio: AI workflow automation for the processes that eat your week, custom AI software development when there is no working product yet, and EU AI Act compliance for systems in scope of Regulation (EU) 2024/1689. The full list is on the services page, and the method is on how an engagement runs.
Trust signals
Proof close to the decision.
Three signals buyers can verify before they start a conversation.
02
Founder-led delivery
The person you contact stays close to the work, from written scope through operation. See how we work.
03
Open-source track record
OpenStore is MIT-licensed and public, with 320 GitHub stars and 120 forks. Inspect the foundation.
Questions
What product and engineering leads ask before starting.
Will this replace our current stack?
No. The feature lands inside the stack you already run: your database, your authentication, your deployment. We add the capability, the evaluation set and the cost ceiling, and we hand back the repository, the runbooks and the operational dashboard. If a proposal ever replaces something that already works for you, we have scoped the wrong job.
How do you keep the AI feature's cost predictable?
Every feature ships with a cost ceiling and per-request tracking from the first week. We measure what a request actually costs, choose the cheapest setup that still passes your evaluation set, and cache what can be cached. When the ceiling is reached, the feature degrades in a way you define in advance instead of producing an unbounded bill.
How do you evaluate an AI feature?
With an evaluation set: real inputs from your data, each paired with the output a good answer would produce. It runs against every prompt, model and retrieval change before release, so a change ships only when it holds the line on your own cases. The set stays in the repository, so the next team can rerun it without us.
What do you need from us to start?
Three things: access to the corpus the feature will read, an evaluation set of real examples, and a place to deploy. Plus a few sentences describing the problem and one metric that says it worked. No roadmap, no org chart and no long discovery phase; the first milestone is scoped and priced before kickoff.
Does our data leave the EU?
EU-first data residency by default. Application runtimes run in the EU, every subprocessor is named on our subprocessors page, and a Data Processing Agreement is published for signature. Personal data stays under GDPR with the processing record you can read. If a specific model provider would move data outside the EU, we say so before you use it.
What does the work cost?
Every engagement starts with a fixed first milestone, scoped and priced before kickoff. The figure is written into the scope together with the success criteria and an explicit out-of-scope list, and it does not move once signed. Ask for the number in your first message and you get it in the reply.
One paragraph is enough to start.
We take on a limited number of contracts each year. If your paragraph names these three things, one reply can tell you whether it’s a fit.
- The problem What is done by hand today, and by whom.
- The metric The one number that will say the AI feature worked.
- The deadline The date it has to work by.