We won the DIMOP Plusz grant: an intelligent voice-based customer service platform

As of today, the project our telecom and backend work of recent years has been preparing us for has officially started: under the DIMOP_PLUSZ-1.1.2/A-24 call we have been awarded funding to develop an intelligent voice-based customer service platform. In this post we summarise the facts of the grant and the planned technical content.

Co-financed by the European Union.

Grant details

• Call: DIMOP_PLUSZ-1.1.2/A-24
• Project identifier: DIMOP_PLUSZ-1.1.2/A-24-2025-00236
• Project title: Development of an intelligent voice-based customer service platform
• Amount of funding: HUF 14,534,144 (14.53 million forints), non-repayable grant
• Funding intensity: 70%
• Implementation period: 1 July 2025 - 14 September 2026

The remaining 30% of own contribution is provided by the company. The project is a significant commitment for a micro-company, which is why we worked out the planning and the milestones in detail over the past months.

Why exactly this?

Over the past years we have worked with SIP telephony, PSTN integration, SMS and WhatsApp channels, and multi-tenant SaaS systems. In every customer support project the same request came up: part of the inbound calls (appointment booking, opening hours, simple status lookups, taking messages) should be handled automatically, without human intervention, around the clock and in several languages.

Advances in speech recognition, large language models and speech synthesis have now made that technically possible. What is missing is a platform that connects these components to the traditional telephone network, and that a small or medium-sized business can configure on its own, without a developer. That is what we want to build.

What will we deliver?

The technical content committed in the project consists of the following main elements.

A multi-tenant SaaS platform. A single infrastructure serves multiple customers, with data, configuration and billing separated per tenant. This builds on the multi-tenant patterns described in our earlier posts: shared schema, row-level security, per-tenant limits.

A no-code flow editor. The customer assembles the course of the call in a visual editor in the browser: greeting, questions, branches, data collection, handover to a human. No programming is needed; the description of the flow is stored in the database, and during the call it drives the assistant.

An STT-LLM-TTS processing chain. Incoming audio is turned into text by continuous speech recognition (speech-to-text), a large language model interprets it and decides on the next step, and speech synthesis (text-to-speech) speaks the answer. Every element of the chain is replaceable, and reducing latency is one of the central technical tasks of the project.

SIP and PSTN integration. The platform receives and initiates calls on real phone numbers over a SIP trunk, building on our own PSTN bridge component. Call audio is forwarded to the processing chain in real time, and back.

Call log and CSV export. For every call, a transcript, a summary and structured data (a booked appointment, for example) are produced, which can be searched in the interface and exported in CSV format into existing systems.

Usage-based pricing. There is no package to pay for up front; the customer pays for the calls actually handled, and billing builds on the usage event model already proven in Abisearch.

GDPR-compliant EU operation. Recordings, transcripts and personal data are stored encrypted on servers inside the European Union, with a configurable retention period, a privacy notice and terms of service.

How do we work on the project?

Development runs the way we are used to: a Docker-based environment, automated tests, staging and production kept separate. Testing voice-based systems is a new challenge, so we also exercise the call flows end to end with real SIP calls, in an automated way, not just with unit tests.

We will report on the project’s progress regularly on this blog: on our experience with reducing latency, on the pitfalls of streaming speech recognition, and on how a language model behaves in the noise of a real phone call. The posts will be written in the usual engineering tone; the grant does not change what we write or how.

We thank the Hungarian Government and the European Union for their support!