AI & Machine Learning Company in Berlin

EU AI Act risk class, more than model complexity, is what decides where a Berlin ML project lands. Engagements run as a scoped AI proof of concept at Berlin rates over six to ten weeks, covering data audit, a baseline model, an EU AI Act risk classification, and a hosted demo, a custom production ML model (fraud scoring, churn prediction, recipe personalisation, sizing models) including MLOps, monitoring, a model card aligned with EU AI Act Article 13, and a Datenschutz-Folgenabschätzung, or a full production AI system with RAG, multiple models, fine-tuning, and SAP or N26-grade integrations. Berlin rates sit above Warsaw and Bucharest because of the BIFOLD, DFKI, and Zalando talent premium but below Zurich and London. We give fixed-fee proposals rather than open T&M, and we invoice in EUR with USD optional for US parents.

The EU AI Act (Regulation 2024/1689) entered force on 1 August 2024, with prohibited-practice rules effective February 2025, GPAI obligations effective August 2025, and high-risk system obligations effective August 2026. The Act classifies systems as prohibited, high-risk (Annex III: employment, credit, education, biometrics, critical infrastructure, law enforcement, migration, justice, democratic processes), limited risk, or minimal risk. Our discovery phase runs the classification on your use case and, for high-risk builds, scopes the conformity assessment, risk-management system (Article 9), data governance (Article 10), technical documentation (Annex IV), record-keeping (Article 12), transparency (Article 13), human oversight (Article 14), accuracy and cybersecurity (Article 15), and post-market monitoring obligations. Even pre-2026, Berlin enterprises (SAP, Zalando, N26, Charité) already require AI Act-aligned governance because the BfDI and Berlin state DPA have telegraphed enforcement priorities clearly.

Yes. For engagements where US LLM providers are off the table — Bundeswehr-adjacent contracts, BSI C5-grade federal procurement, parts of the Charité research portfolio, mittelstand IP-protection-first deployments — we default to Aleph Alpha’s Luminous and Pharia models hosted on German infrastructure, T-Systems Sovereign Cloud (operated by Deutsche Telekom inside Germany), Open Telekom Cloud, or IONOS Cloud. DeepL handles translation-heavy workflows where its German-language data moat outperforms US frontier models. Where a hybrid stack is acceptable, we route only non-sensitive prompts to OpenAI or Anthropic through Azure Germany and keep model fine-tuning, embeddings, and RAG context inside German operator boundaries. The architecture decision lands in week one with the DPO and CISO in the room, not as an afterthought.

GDPR provides the EU baseline, BDSG layers German-specific obligations (notably §26 BDSG for employment data and §22 BDSG for special categories), and TTDSG governs cookies and telemedia data that any client-side AI surface touches. Every engagement opens with a Datenschutz-Folgenabschätzung (Article 35 GDPR DPIA) where personal data is in scope, a §26 BDSG check whenever employment use cases (CV screening, performance prediction) are involved, and a Betriebsrat (works council) coordination plan under §87 BetrVG where AI-supported employee monitoring is implicated. TTDSG compliance is built into the front-end so consent banners, fingerprinting risk, and on-page personalisation all survive a BlnBDI audit. Records of processing activities (Verzeichnis von Verarbeitungstätigkeiten) are produced for the DPO. Cross-border transfers are governed by Standard Contractual Clauses plus a Transfer Impact Assessment, with Schrems II considerations documented explicitly.

When a project genuinely requires novel research (medical imaging architectures, rare-event modelling, federated learning across hospital networks, sovereign-LLM fine-tuning), we scope collaborations with BIFOLD-affiliated faculty at TU Berlin, DFKI Berlin researchers, or Charité’s AI-aligned clinical labs rather than overselling in-house capability. Our core team owns applied engineering, MLOps, and productionisation — which is where most Berlin AI projects actually stall. For standard work (RAG, fine-tuning, classical ML, computer vision on known architectures) no academic partner is needed. We will tell you up front which bucket your problem falls into, and we have brokered industry-funded research arrangements through DFKI’s technology transfer office when a roadmap genuinely warrants it.

We default to AWS eu-central-1 (Frankfurt), AWS eu-central-2 (Zurich) for redundancy, Azure Germany West Central (Frankfurt) for SAP-integrated workloads, and GCP europe-west3 (Frankfurt) for storage, training, and inference. For BSI C5-grade or sovereignty-first engagements we deploy on T-Systems Sovereign Cloud (Deutsche Telekom-operated, German jurisdiction), Open Telekom Cloud, or IONOS Cloud. LLM layers that must remain inside Germany use Aleph Alpha Pharia hosted on those operators or self-hosted Llama 3 and Mistral on EU GPU instances. Non-EU transfers are permitted only after a Transfer Impact Assessment under Schrems II, typically reserved for non-personal-data internal tooling. Residency is documented in the model card and the Auftragsverarbeitungsvertrag (Article 28 GDPR data processing agreement) so the BfDI, BlnBDI, and your DPO have a single defensible answer.

A typical Berlin mittelstand model (industrial demand forecasting, predictive maintenance, document extraction, sales-funnel scoring) takes twelve to twenty weeks from kickoff to production, assuming clean historical data and EU AI Act Annex IV documentation as part of scope. Weeks one to four are data audit, baseline modelling, DPIA, and AI Act classification. Weeks five to twelve are iteration, challenger testing, and bias review. Weeks thirteen to eighteen are MLOps, monitoring, shadow-mode deployment, Betriebsrat sign-off where applicable, and DPO review. Week nineteen onward is gradual rollout. If you are pre-data or need labelling, add six to eight weeks. For Charité-grade clinical AI under MDR, add a twelve to twenty week regulatory pathway with a notified body. We have shipped to this cadence inside SAP-ecosystem and BaFin-regulated stacks.

Most custom AI work qualifies for the Forschungszulagengesetz (Research Allowance Act), Germany’s federal R&D tax credit. The programme provides a 25 percent refundable credit on eligible personnel costs up to EUR 10M per company per year (raised from EUR 4M in 2024), and applies whether the work is done in-house or contracted to qualified service providers — including our team. Eligible activity covers novel modelling, architecture experimentation, algorithmic uncertainty, and systematic investigation, but not routine integration. We deliver time-tracked logs, technical narratives, experiment histories, and failed-hypothesis documentation aligned with the BSFZ (Bescheinigungsstelle Forschungszulage) certification process. Berlin clients can often combine Forschungszulage with Investitionsbank Berlin programmes for further support. Final eligibility sits with your Steuerberater and the BSFZ.