Services · 03
AI & data science
We bring computer vision, machine learning and language models into existing processes. We test our methods in our own lab, and we document every system for the EU AI Act.
The problem
The prototype is quick to build. Then it gets serious.
A demo chatbot or a first classification model is often running within a few days. The real work starts after that. How good is the data really, who is allowed to use it, how do you notice when the model is wrong, and what happens in the process when it is?
On top of that, the EU AI Act has been in force since August 2024, with its obligations applying in stages. The prohibitions have applied since February 2025 and the transparency obligations since August 2026. Regulation (EU) 2026/1744 (the Digital Omnibus) postponed the obligations for high-risk systems to 2 December 2027, and to 2 August 2028 for AI in regulated products. Depending on the purpose, the obligations range from transparency notices to full risk management.
What you get
Deliverables
- AI use-case check: feasibility, data readiness and EU AI Act classification for one specific use case
- Data preparation and analysis: merging, cleaning and standardising data from several sources
- Models for forecasting, classification and computer vision, including multi-camera setups
- Language models in business applications, with a test set that includes wrong, ambiguous and unanswerable questions
- Documentation for the EU AI Act: purpose, data provenance, tests, known limitations, human oversight
Technologies from our AI projects
- Data and ML
- Python, pandas, scikit-learn, TensorFlow
- Computer vision
- Convolutional neural networks combined with classical image processing
- Optical inspection
- Multi-camera setups
- Language models
- Open-weight models such as Qwen, run locally with Ollama; reinforcement learning training (PPO) in the lab
Approach
Measure first, then build
- We agree how success will be measured and build the test set before any model exists.
- We check the data: origin, quality, legal basis. Personal data from your project stays in the EU, and at Senva only the Berlin team has access to it.
- We compare simple methods with complex ones. If a rule or a classical model does the job, we say so.
- We document what the system is meant for, and what it is not meant for.
- We hand over with measured results and a plan for monitoring the model in production.
What the lab contributes
In Senva Labs we study how prompt wording shifts results and whether language models stick to rules. We use the methods from that work, such as control conditions and fixed seeds, for the tests in your project as well.
Work
AI and data projects
-
Visual drone detection pilot
Read the case studyFor Deutsche Telekom's drone shield programme, 2017. A CNN combined with classical image recognition.
-
Optical inspection of circuit boards
Read the case studySeveral cameras, ML-based image processing, Python. March to December 2018.
-
Education data and performance forecasts
Read the case studyData preparation and forecasting models for an education institute in Vietnam, 2019. pandas, scikit-learn, TensorFlow.
Typical first step
AI use-case check
Afterwards you know whether your use case is feasible with your data, which obligations follow from the EU AI Act and roughly what a prototype would cost. We agree scope, duration and a fixed price in writing before we start.
FAQ
Questions from CTOs and data protection officers
Does our use case fall under the EU AI Act?
Formally, almost every use of AI does. What matters is the risk category. Many internal applications are subject only to transparency obligations, or to no specific obligations at all. High risk covers, among other things, systems for recruitment, creditworthiness assessment and education, and safety components of critical infrastructure. Under the Digital Omnibus, their obligations apply from 2 December 2027. In the AI use-case check we classify your case and write down the reasoning. This does not replace legal advice.
Can we use language models without personal data leaving the EU?
Yes, either with models that a cloud provider offers in an EU region, or with open-weight models on your own infrastructure. For cloud models we check with you whether the provider's contract terms actually limit processing to the EU. Which option fits depends on the type of data, the load and the budget, and we'll show you the numbers.
Who at Senva sees our training data?
If it contains personal data, only the Berlin team sees it, and only within the scope of the data processing agreement (AVV). The partner team in Hanoi works only with synthetic or anonymised data. We treat pseudonymised data as personal data.
What does your lab bring to our project?
We have trained language models ourselves, tested them against rules, found bugs in our own code along the way and corrected them in public. We bring those habits into your acceptance criteria: control conditions, fixed seeds and a list of what a result does not prove.
Intro call
Bring a specific use case
In 30 minutes we'll work out whether an AI use-case check is worth doing for your case. You'll be speaking with Kim Pham.
- info@senva.de
- Phone
- +49 30 5268 7541
- Address
- Bundesallee 32, 10717 Berlin, Germany