AI product engineering
We build AI features that hold up outside a demo: retrieval that cites its sources, evaluations that catch regressions before customers do, and a cost model you can defend to a CFO.
Most AI projects fail after the prototype, not during it. The demo works, then it meets real documents, real latency budgets and real auditors, and nobody can explain why an answer changed between Tuesday and Thursday.
We build the unglamorous half: retrieval with provenance, evaluation harnesses wired into CI, guardrails and fallbacks across providers, and token and latency budgets tracked per feature.
On DoldAdress.se we replaced a manual email review team with a classifier: Google answers every delisting request by email, and a model now decides the outcome so a person only sees the ambiguous cases. Sodoko is the same discipline applied to our own product — an agent that writes production Odoo modules and has to be right, not plausible.
What you get
- A retrieval or agent system in production with evaluations your team owns
- Provider-agnostic routing, so a model deprecation is a config change
- Cost and latency instrumented per feature, not per month
- Data handling documented for GDPR and internal review
Typical stack
- OpenAI
- Evaluations
- RAG
- AWS Bedrock
Where we run this ourselves
- Sodoko — ai that ships production code
- DoldAdress.se — taking the manual work out of a privacy service
Start here
A one-week discovery gives you a written plan and a cost, whether or not you carry on with us.
Start a project