A self-hosted model on our own GPU
Our main tool is a 27-billion-parameter language model running on our own GPU with 32 GB of memory. It serves internal services around the clock: document parsing, call summarisation, content generation, development assistance. The context window is 128,000 tokens — the model holds a hundred-page document in memory at once.
The practical takeaway we pass on to clients: a single consumer-class GPU is enough to cover the work of a small company. The “you need a cluster” conversation starts when the user count runs into the hundreds.
Telephony: 2,254 calls through local transcription
Between March and August 2026 our system handled 2,254 calls. Each one is transcribed by a local speech-recognition model with speaker diarisation, after which a local language model produces a summary: what was discussed, what was promised, whether a response is needed. Recordings and transcripts are personal data, so they are encrypted at rest and kept under a retention policy.
Antifraud: 60% of traffic turned out to be bots
Our protection runs on five sites, including our own. Events processed and verdicts issued:
| Metric | Value |
|---|---|
| Events processed | 108,789 |
| Classified as bots | 65,853 |
| Share of bot traffic | 60% |
| Sites protected | 5 |
Six events out of ten are automation rather than people. We do not present this as an industry average: these are our sites and our traffic, and your share will differ. But the order of magnitude explains why ad budgets drain unnoticed — see what click fraud is.
Monitoring: 58 nodes, one alert channel
A watchdog of our own looks after the infrastructure: 58 nodes — services, disks, GPU memory, queues, certificates, backup freshness. The checks go beyond “the process is alive”: a synthetic request goes to the language model periodically, because a process can be running and still have stopped answering.
- •Backups are verified by restoring them, not by the existence of a file,
- •Alerts go to a single channel — otherwise people stop reading them,
- •Every incident ends with a root-cause review rather than a service restart.
What this means for you
We do not sell what we have not tested on ourselves. If we propose putting a model inside your perimeter, it is because we live without external APIs. If we say bots consume a noticeable share of traffic, it is because we see it on our own counters. Any number on this page can be walked through in a live demonstration.
How a server for this is built and what it costs is on the on-premise AI server page, and how to calculate the payback is in how to adopt AI in your business.