Real work

AI systems already running in production

We do not show demos. These are real projects built by TenBeltz, operating with real traffic, real users, and real costs.

Client · IrontecTelecommunications · Contact center

AI-powered call analysis for contact centers

Irontec, a leading technology consultancy in telecommunications and open source, hired TenBeltz to design and build an enterprise call-analysis platform.

What we built

Transcription with multiple speech engines (Deepgram, Whisper), conversation analysis with LLMs — service quality, data extraction, metrics — RAG over the calls, asynchronous processing with workers, and deployment on Kubernetes. Model selection and on-premise deployment driven by privacy requirements.

TenBeltz has operated as Irontec’s AI engineering partner since 2025.

Client · QamareroHospitality · SaaS

Voice agents handling real bookings

Qamarero, a hospitality management SaaS, wanted to bring AI into its product without breaking the daily operation of its restaurants.

What we built

A system of phone agents that answers calls and manages reservations end to end, integrated with the CRM and the SaaS’s own table and schedule management. Several months of joint work until it was running in production.

Conversational voice AI wired into the real operation, not an isolated chatbot.

Legal sectorLegal · Expert reports · client under NDA

Multi-agent system for expert legal reports

An expert report is a legal document: a wrong date or a misassigned injury has real consequences. Here, precision rules.

What we built

Specialized agents per report section, extraction into validated structured data, and a double anti-hallucination layer: deterministic rules (dates, inconsistencies against the medical assessment scale) and automatic evaluation against real reports, section by section. Multi-provider, routed per task.

Result

−42%

cost per report with no loss of precision.

You do not cut costs by picking the cheap model. You cut them when your evaluation system proves the cheap one breaks nothing: on this system that meant −42% cost per report with no loss of precision.

Document managementDocument processing · client under NDA

Automatic classification of thousands of documents

Classifying documents into 15 categories and 77 subcategories. The question was not which LLM to use, but whether one was needed at all.

What we built

A measured comparison of three approaches on real data: semantic with embeddings, 100% LLM, and hybrid. The hybrid reaches 95.5% top-5 accuracy with 17% fewer LLM calls.

Measure before assuming you need an LLM.

Capabilities

What kind of systems we build

Voice agents connected to the real operation

RAG and document classification at scale

Multi-agent systems with automatic quality evaluation

On-premise AI in regulated environments (ENS, EU AI Act)

If the project matters, it is worth defining it properly from the start

Request proposal

Tell usyour project

Share the context and we will assess whether there is a strong fit for TenBeltz.

hello@tenbeltz.com

+34 640 520 819

Prefer to talk it through?

Book 30 min with Aritz