See what your AI actually does, before your customers do.

We trace every prompt, answer and euro your AI spends. Quality, cost and compliance on one screen, kept in the EU.

AI is in production. Most teams cannot see inside it.

Most large companies run several AI applications, built by different teams on different models, with their own logs or none. When an answer goes wrong, nobody knows if it is a one-off or a pattern, which prompt caused it or what it cost.

Standard application monitoring misses this. A model call can succeed and still be wrong, off-policy or ten times more expensive than last week. Quality and cost must be measured per request. Open standards such as OpenTelemetry and tools such as Langfuse now make that practical.

Finance wants spend per business unit. Risk and works councils ask what the system says and who checked it. Under the EU AI Act, high-risk systems need automatic logs that deployers keep. Without this evidence, projects stall at pilot stage or get switched off.

What we build

  • End-to-end tracing

    Every prompt, answer, retrieval and tool call becomes one trace, in Langfuse, Arize Phoenix or your existing monitoring.

  • Quality, latency and spend dashboards

    Quality, response times, tokens and euros spent, per business unit, application and model, updated continuously, not in monthly exports.

  • Evaluations and regression tests

    Tests from real cases and your experts' judgements. A prompt change or model upgrade must pass before it reaches users.

  • Drift and hallucination alerts

    Live answers scored for grounding, policy breaches and user dissatisfaction. When scores drop, the owner is alerted, traces attached.

  • AI gateway and cost control

    One gateway in front of all model providers sets team budgets and sends simple requests to cheaper models.

  • Review queues for domain experts

    Flagged conversations reach your service, legal or medical experts, no code needed. Their verdicts feed back into tests and prompts.

How it works

One request, from a user's question to someone acting on a problem.

  1. Request passes the gateway

    The gateway checks team budget and permissions, redacts sensitive data, picks the model and records cost and latency.

  2. Every step is traced

    One trace per request, tagged by team and use case, stored in the EU or your own cloud.

  3. Answers are scored

    LLM-as-a-judge checks and fixed rules score each answer for grounding, compliance and quality. Scores stay on the trace.

  4. Problems reach an owner

    Threshold breaches trigger alerts. Flagged sessions go to experts, who rate severity and mark what should change.

  5. Fixes are tested before release

    Reviewed examples become tests every prompt or model change must pass, so old problems cannot quietly return.

How we roll it out

  1. Baseline

    2 to 3 weeks

    We list your AI applications, owners and models, then agree metrics, retention and hosting with IT, risk and data protection.

    • Inventory of AI applications and model calls
    • Target metrics and alert thresholds
    • Hosting and data protection design
  2. Instrument

    4 to 6 weeks

    Tracing and gateway go into your EU environment. Two or three live applications get dashboards, cost allocation and first evaluators.

    • Tracing and gateway running in your environment
    • Dashboards per team and use case
    • First evaluators and alerts on live traffic
  3. Operate and extend

    Ongoing

    We add applications and build test sets with your experts. Your team takes over, and we stay until targets hold.

    • Regression suites in CI/CD
    • Review queues staffed by your experts
    • Runbooks and handover to your platform team

In production elsewhere

Public examples from other organisations, not unus clients, showing what AI observability looks like when done properly.

Public case studyMerck GroupPharmaceuticals, life science and electronics · Germany · 2025 One observability platform for AI teams across the group Approximately 80 use cases on one platform
The situation
Merck's central AI team serves all three business sectors and functions from R&D to manufacturing and safety. With more than 300 generative AI use cases, it needed one view of every application, without a separate tool per project.
What they deployed
Merck chose Langfuse after a structured market evaluation, partly because self-hosting keeps sensitive pharmaceutical data in-house. Approved projects are set up automatically. Traced applications include an R&D lab assistant over more than 80,000 documents and the myGPT suite.
The outcome
Approximately 80 of Merck's more than 300 generative AI use cases run on Langfuse, with over 200 people working on it regularly. The case study reports development cycles cut from weeks to days and audit trails of every AI interaction for governance.

Treat observability as a shared, self-hosted platform with automated onboarding, not a tool each project picks for itself.

Source: Langfuse, How Merck is serving 80+ AI project teams globally on Langfuse, September 2025
Public case studyDKB (Deutsche Kreditbank)Banking · Germany · 2026 Every banking assistant answer screened and reviewed Screening every answer, before every release
The situation
DKB's in-app assistant, live since spring 2024, serves around 6 million customers and handles 10,000 to 20,000 conversations a day. In 2025 service experts worked from CSV and Excel exports, so isolated errors were hard to tell from patterns.
What they deployed
DKB self-hosts Langfuse, as customer data could not go to an outside vendor. An LLM-as-a-judge screens every answer for compliance issues such as investment advice, grounding and customer satisfaction. Specialists rate flagged sessions. Under-18s and employees are not screened automatically.
The outcome
Reviewed examples become datasets that run as regression tests before every release. DKB describes screening as part of its licence to operate: a senior machine learning scientist there says that if screening were disabled, all agent activity would have to stop.

In regulated industries, continuous screening and expert review are what allow an AI agent to stay in production at all.

Source: Langfuse, How DKB uses Langfuse to keep its AI banking assistant accurate, compliant, and trustworthy at scale, August 2026
Public case studyZalandoE-commerce · Germany · 2026 A central LLM gateway with built-in cost tracking 2k monthly active users on six small pods
The situation
Zalando has more than 250 engineering teams and never mandated a single AI tool. Beyond GitHub Copilot, engineers wanted direct API access to several model providers for self-built agents, which raised questions of authentication, cost, version control and measuring adoption.
What they deployed
In January 2024 the ML platform team deployed a LiteLLM-based proxy to OpenAI, AWS Bedrock and Google Vertex. Hooks track costs anonymously, enforce client upgrades and retire old models. Prompt caching is added automatically, and a local tool shows costs.
The outcome
The proxy serves 2k monthly active users on six small pods with 2 CPU cores and 4 GB memory each, and gives the platform team a single point to measure adoption. Zalando states the automatic prompt caching reduced costs for custom agents.

A thin gateway in front of every model provider gives you cost tracking, model governance and vendor choice without slowing teams down.

Source: Zalando Engineering Blog, Agentic Engineering at Zalando: a snapshot, August 2026

These are public deployments by other organisations, summarised from the sources linked. They are not unus clients.

Built for European rules

Traces stay in the EU

Everything runs self-hosted in your cloud or EU or Swiss data centres. Prompts and answers, often holding customer data, stay there.

GDPR and nDSG by design

Personal data masked or pseudonymised before storage, retention per application, role-based access. Employees or minors can be excluded from automated review.

EU AI Act record-keeping

Annex III high-risk duties start 2 December 2027: automatic logs, kept at least six months. Our traces provide them and prove oversight.

Human oversight you can prove

Every flag, review and release test is logged with who, when and why. Auditors and works councils see a trail, not anecdotes.

Questions buyers ask

We already use Datadog or Dynatrace. Why do we need anything else?

Infrastructure monitoring tells you a call succeeded and how long it took. It does not tell you whether the answer was correct, grounded or compliant, or which team's use case spent the money. We send GenAI traces in OpenTelemetry format, so they can sit next to your existing monitoring, and add the evaluation, review and cost layers it lacks.

Can we keep prompts and answers out of a US vendor's cloud?

Yes. We prefer open-source tools such as Langfuse and LiteLLM that can be self-hosted on your own cloud or in an EU or Swiss data centre. Prompts, answers and scores stay in your environment. Where you choose a managed service, we select an EU region and document the data flows for your data protection officer.

Does logging every conversation create a GDPR problem of its own?

It can, which is why we design for it. We mask or pseudonymise personal data before storage, set retention periods per application, restrict access by role and exclude groups you are not allowed to review automatically. Your data protection officer signs off the design in the baseline phase, before any production data is captured.

How do you detect hallucinations automatically?

No single check catches everything. We combine grounding checks that compare answers with the retrieved sources, LLM-as-a-judge evaluators tuned on your experts' judgements, and deterministic rules for things that must never happen, such as giving investment advice. Anything uncertain goes to a human review queue, and confirmed cases become regression tests.

Will a gateway add latency or lock us into one provider?

We measure the gateway's overhead on your own traffic during rollout, so you can judge it against model response times before going further. On lock-in, it works the other way: applications call one interface, and you can switch or mix providers, including open-weight models on your own infrastructure, without changing application code.

Do we have to rebuild our existing AI applications?

No. Most frameworks have OpenTelemetry or Langfuse integrations, so instrumentation is usually a matter of adding a library and a few tags. Routing calls through the gateway is a configuration change. We start with two or three production applications and extend from there.

Bring one process. We'll bring the loop.