Tracebloc's EIC Accelerator Trains AI on Data It Never Sees

Two problems meet in the same place: enterprises hold data they are not allowed to share, and they buy AI models they cannot test on it. Tracebloc, founded in Berlin in 2020 by Lukas Wuttke, sends the model to the data. A company deploys a tracebloc workspace on its own infrastructure, defines a task and a dataset, and invites outside contributors, who submit models, see only metadata and never touch the raw records. Training and evaluation run inside the client's environment, and the contributors get metrics back.

The EIC Accelerator selected the company at its October 2024 cut-off under the challenge for human-centric generative AI made in Europe, as a grant-only award, and CORDIS records €2,348,290.71 from June 2025 to May 2027 to build what the project calls a global data access layer for AI, with benchmarking of models for accuracy, compliance and energy efficiency. Seventeen months in, no customer, no revenue and no periodic report are public.

The model goes to the data

The platform is a commercial federated learning system: each client trains locally on its own data, only model updates go back to a central server for aggregation, and tracebloc runs the orchestration, parallelisation, scaling and monitoring. A workspace can be deployed on a laptop, a Linux or Windows machine or a GPU server, and contributors submit models through an SDK from a notebook.

The company pitches it at five concrete frustrations, which its homepage spells out: a vendor claiming 95% accuracy on a demo set the buyer cannot reproduce; external collaborators blocked by NDAs and security reviews; regulators asking for evidence that a model was tested before deployment; fine-tuning small language models on clinical notes or legal contracts that cannot leave the building; and reproducing a paper's results in an identical container rather than a different software stack. Five vendors, it argues, can submit to the same workspace and be judged on the same holdout set.

What is public: templates, not customers

The company's hub lists 15 public use case templates with leaderboards, each a task on private data: mammography screening across five sites, prostate cancer radiation dose optimisation, heart disease risk prediction, diabetic retinopathy validation across clinical sites, fine-tuning language models on medical abstracts, rare disease biomarker panels, a six-site multi-omics study, drug-induced liver injury detection, paediatric bleeding disorder stratification under the EU AI Act, warranty claims classification and weld defect inspection in EV manufacturing, Sentinel-2 crop classification, financial time series forecasting and drone object detection for smart cities.

These are templates on the platform with scores from named contributors, not disclosed deployments. No customer company is named anywhere on the site, and the company has published no revenue, user numbers or case study with a client. Its blog, written mostly by the founder, argues the case for federated learning in healthcare, drug discovery data access, LLM benchmarking and AI vendor evaluation.

1. Working federated platformDeployableInstall commands, docs and SDK are public.
2. Benchmarking for compliance and energy useNot publishedNamed in the project objective; no results public.
3. Enterprise customersNone namedOnly public templates, no client case study.
4. Project reportingNothing yetCORDIS page unchanged since June 2025.

A grant and nothing else disclosed

Tracebloc's award is grant only: the EIC gave €2.35 million and took no equity, which also means no EIC Fund due diligence on the company's valuation. No venture round has been announced, no investor is named on the company's website, and its German accounts were not obtained for this article. For a software company with a product already deployable, that leaves the EU grant as the only disclosed source of capital for the current phase.

Compliance as the selling point

The pitch is built on European rules. The project summary names the EU AI Act, and one of the hub's own templates is labelled as an EU AI Act compliance case. Keeping data inside the company and shipping only model updates is an answer to the data protection and intellectual property problems that stop European hospitals, manufacturers and banks from working with outside AI teams. That is also why the EIC put this project under its human-centric generative AI challenge rather than treating it as infrastructure.