Collaborate with MagCIL
We work with companies, public bodies and research consortia on problems where audio, speech, music, image or video analysis is part of the answer. Two routes are described below: contract work with industry and the public sector, and partnership in collaborative research projects. Both are contracted through the Institute of Informatics and Telecommunications of NCSR "Demokritos".
Industry and the public sector
Engagements are usually bilateral contract research or consultancy with NCSR "Demokritos", scoped so that each stage produces something you can act on. We normally start small and only grow the scope once a problem has been shown to be tractable on your own data.
What we undertake
| Feasibility studies | A short, bounded engagement that tests whether your problem is solvable with today’s methods and the data you actually have. Ends in a written assessment: what is achievable at what accuracy, what the data would need to look like, what it would cost to build, and where the approach is likely to fail. |
| Custom model development | Training and adaptation of models for your domain and your recording conditions — speech, music, audio events, images and video — including the low-resource cases we work on routinely: few-shot classes, little or no labelled data, Greek and other under-resourced languages. |
| Evaluation and benchmarking | Independent evaluation of systems, yours or a vendor’s: evaluation protocol and dataset design, annotation, measurement under noise and distribution shift, and adversarial robustness testing. Useful before you procure a system, and before you trust a number in a datasheet. |
| Prototype integration | A working prototype that runs in your environment rather than in a notebook: containerised inference bundles for on-premises or private-cloud deployment, edge and real-time variants where the compute budget demands it, and the packaging needed to hand the result over to your engineers. Our deepaudio-lab platform is often the starting point for audio work. |
What to bring to a first conversation
- The problem. Stated as the decision, product or process you need to improve, rather than as a technology to apply. If there is a current solution, manual or automated, tell us how well it performs and what "good enough" would mean.
- Data availability. What recordings or images exist today, how many hours or items, how they were captured, how much of the material is labelled and by whom, and under what terms we may use it — ownership, licensing, personal data and GDPR basis. Where no data exists yet, say so: designing the collection is often the first piece of work.
- Deployment constraints. Where the model has to run (cloud, your own servers, an edge device), the latency and throughput you need, the cost per hour of input you can carry, privacy limits on where data may travel, and whether the use case is regulated — the EU AI Act in particular changes what we would propose.
Research consortia
We take part in national and European projects as a technical partner, and we are selective: we join where a work package needs the methods we actually work on, and we are happy to say when a call is a poor fit. Recent and ongoing examples include the PHAROS AI Factory, DeployAI, MI-TRAP and FaRADAI — see Projects for the full list.
Roles we can take in a consortium
| Audio-analysis components | Speech analytics (emotion, paralinguistics, diarization, speech recognition for Greek and other low-resource languages), audio event detection and acoustic scene analysis, soundscape monitoring, music information retrieval and audio fingerprinting. Delivered as components with defined interfaces, not as prototypes left in a repository. |
| Multimodal modelling | Fusion of audio, vision and text; self-supervised and few-shot representation learning; generative modelling for audio, music and images. Typically the technical core of a WP rather than a contribution to one. |
| Robustness and evaluation | Adversarial robustness of audio and vision models, benchmarking under noise and distribution shift, synthetic-media and deepfake detection, and independent evaluation of partners’ components against protocols we design with them. |
| Data, annotation and benchmarks | Dataset design and collection, semi-automatic and AI-assisted annotation pipelines, and the release of datasets and benchmarks — including for repertoires and languages that mainstream models represent poorly. |
| Dissemination, exploitation and training | Open-source releases through our GitHub organisation, public-facing demonstrators and AI-literacy platforms, and training through the MSc in Artificial Intelligence we co-organise (see Education). |
Partner profile
A one-page profile of the group, suitable for attaching to a proposal or circulating inside a consortium: expertise, infrastructure, project track record and legal-entity details.
Contact
For both routes, write to the group’s scientific contact with a short description of the problem or the call; we will tell you quickly whether we are the right partner.
Dr. Thodoris Giannakopoulos
Director of Research, Head of the Multimodal Analysis Group (MagCIL)
Institute of Informatics and Telecommunications
National Centre for Scientific Research "Demokritos"
Patriarchou Grigoriou E’ & 27 Neapoleos St., 15341 Agia Paraskevi, Athens, Greece
tyianak@iit.demokritos.gr
For proposals with a deadline, include the call identifier and the submission date in the first message.