NannyML's EIC Accelerator Grant Came Days Before the Soda Sale

An AI model that works on launch day can quietly get worse. Customers change, prices move, sensors drift, and the model keeps making decisions on a world that no longer exists. The usual check, comparing predictions with what actually happened, often has to wait weeks or months for the real outcome. NannyML, a Leuven start-up founded in 2020, built an open-source Python library that estimates a model's performance before those outcomes arrive.

The EIC Accelerator selected NannyML at its October 2024 cut-off for blended finance, and the Commission signed a grant of €2,043,691.13 for the project EU-AURA on 29 May 2025, according to CORDIS. Eleven days later, on 9 June 2025, the Brussels data quality company Soda announced that it had acquired NannyML. The EIC project, meant to bring the technology to market, now runs inside a company that has become part of someone else's product.

Estimating accuracy without the answers

Most monitoring tools watch the data going into a model and raise an alarm when it drifts. NannyML went one step further: its library estimates the model's actual performance in production without access to the true labels. For classification models it uses the model's own confidence scores to estimate accuracy, a method the company calls confidence-based performance estimation; for regression it trains a second model to predict the errors of the first, which it calls direct loss estimation. The library on GitHub is published under the Apache 2.0 licence and covers tabular classification and regression models.

The open-source project drew a real following. It has 2,160 stars and 210 forks on GitHub, and the CORDIS summary of the EIC project cites more than 2,000 monthly active instances, 20,000 monthly downloads and a community of 50,000 data scientists, with EssilorLuxottica, Ahold Delhaize and TUI among its users. Those usage figures come from the company's own proposal and are not published elsewhere.

NannyML's commercial product was a hosted version for companies. In July 2023 the fintech software house Ailleron announced a partnership to integrate NannyML's monitoring into its models for banks, pitching it as a way to keep models compliant with the EU AI Act.

A pre-seed round, then four years of losses

NannyML raised a €1.06 million pre-seed round in October 2020, led by Volta Ventures with Lunar Ventures and angels including Stijn Christiaens of Collibra and Jonathan Cornelissen of DataCamp, according to the Start it @KBC announcement. The company announced no later round.

The accounts filed with the National Bank of Belgium show what followed. NannyML lost between €0.54 million and €0.83 million a year from 2021 to 2024, and equity fell from €1.02 million at the end of 2020 to minus €1.69 million at the end of 2024, with €1.77 million of debts against €70,684 of assets, according to Staatsbladmonitor and Companyweb. By the time the EIC grant was signed, the company was funding itself with debt.

Sold eleven days after the signature

Soda announced the deal on 9 June 2025 as the first day of its launch week, and Silicon Canals and Signalbase reported it the same day, without financial terms. Soda, founded in 2018 by Maarten Masschelein and Tom Baeyens, said NannyML's team and technology were now integrated into its platform and that its algorithms would reduce noise in Soda's data quality checks. Both Soda and NannyML's newsletter promised that the open-source project "will remain open, maintained, and fully supported".

The Belgian register shows the reshuffle. The Official Gazette of 7 July 2025 records changes to the articles of association, the capital and shares, and resignations and appointments; since 20 June 2025 Hakim Elakhrass is the sole director listed in the company register, and the share capital stands at €1,331,667.08. The register does not list shareholders.

The first accounts after the sale cover 15 months to 31 March 2026. They show an operating profit of €558,757, equity back at €515,451 and debts down to €146,250, with one full-time employee. The accounts do not disclose turnover, so they do not show whether the profit came from the EIC grant, from Soda or from customers.

An EIC project inside an acquired company

EU-AURA, short for European Union AI Uncertainty Reduction and Alignment, was to develop five algorithms that estimate a model's performance and detect failures and bias after deployment, and to commercialise NannyML's technology. CORDIS lists the project from 1 June 2025 to 31 May 2027 with NannyML of Leuven as coordinator, and the page was last updated in March 2026. It shows no periodic report, result or publication yet.

The blended award also included an equity offer from the EIC Fund. Whether that investment was made is not public: the EIC Fund does not list NannyML among its portfolio companies in any source found for this article, and neither Soda nor NannyML has mentioned it.

The open-source record is the most visible measure of the promise to keep the library maintained. The last release, version 0.13.1, was published on 12 July 2025, and there has been no commit since, according to the GitHub repository. Fifteen months of silence does not end a project, but it is not what "fully supported" usually means.

1. Five performance and bias estimation algorithmsNot reportedNo result or publication on CORDIS.
2. Commercialisation of NannyML's technologyAbsorbedSoda says the technology is integrated into its platform; no NannyML product launch is public.
3. Maintained open-source libraryStalledLast release and commit on 12 July 2025.
4. Project completion by May 2027Too earlyThe project is still running.

Europe's AI Act needs exactly this kind of tool

The EU AI Act requires providers of high-risk AI systems to monitor them after they go into service. Tools that tell a bank or a hospital when a model has started to fail, before the damage shows up in the outcomes, serve that obligation directly, which is why the EIC funded EU-AURA. The technology stays in Europe: Soda is a Brussels company, and the EIC project remains with the Leuven entity. Whether the EU money produces the promised algorithms, or simply supports a team that now works on Soda's roadmap, will show in the project's reports.