Use case · Infrastructure, operations, AI

Anonymised, like all our use cases

Voice AI without a cloud provider.

For the AI call analysis on our service platform, we operate our own AI infrastructure: speech recognition, word timestamps and the language model run on a GPU server in a data centre in Germany. No call leaves our systems.

Deployed in

AI call analysis

Services

Infrastructure, operations, optimisation

Location

Data centre in Germany

Hardware

1 GPU server, 20 GB GPU memory

Since

2026

Measured, not estimated.

Since the timestamp service moved to the GPU, analysing a three-minute call takes about 47 seconds instead of 80. Each service has a fixed share of the GPU memory.

How the platform is built.

  1. Gateway All requests pass through a gateway, encrypted and only after login.
  2. Speech recognition A speech recognition model transcribes every call.
  3. Timestamps A second model adds a timestamp to every word, directly on the GPU.
  4. Language model A local language model assigns sentences to speakers, summarises and detects signals.
  5. Operations All services run in containers, with monitoring and a rollback for every change.

≈ 47 s

Analysis of a 3-minute call, previously about 80 seconds

14×

faster: word timestamps in 2.1 instead of 29.7 seconds

20 GB

GPU memory, split into fixed shares for three AI services

0

external AI services, all data stays in Germany

What makes the difference.

Data stays in Germany

Recordings and results do not leave our infrastructure. No cloud provider listens in.

Measure, don’t guess

We back every optimisation with measurements for each processing step, before and after.

Resources under control

GPU memory is planned per service. We know how much headroom is left.

Rollback built in

Every change has a documented rollback. That keeps operations stable.

Certified data protection and security

Data protection officer (IHK)

Information security officer (TÜV)

Achim Groeling, Managing Director

What we built it with

  • Linux
  • Docker Compose
  • NGINX
  • NVIDIA GPU (20 GB)
  • CUDA
  • vLLM
  • llama.cpp
  • Qwen3-ASR
  • Qwen3-ForcedAligner
  • FastAPI

Where should your AI run?

In an intro call we work out whether our servers, a data centre of your choice or your own hardware is the right fit.

Book an intro call