AI & Machine Learning Company in Hamburg

We give fixed-fee proposals instead of publishing bands or open T und M estimates. A scoped AI proof of concept is a six to ten week engagement covering a data audit, a baseline model, and a hosted demo on eu-central-1 or Azure Germany West Central. A production ML model (route optimisation, recommendation, demand forecasting, claims triage, dermatology vision baseline) adds MLOps, monitoring, and a DSGVO-aligned model card. Full production AI platforms layer on RAG, fine-tuning, multiple models, and SAP, Hapag-Lloyd, or Otto Group enterprise integrations. Hamburg sits roughly level with Berlin and Düsseldorf but below Munich, and we give fixed-fee proposals instead of open T und M estimates.

Yes. We have scoped vessel routing, bunker-fuel optimisation, reefer-container temperature anomaly detection, ETA prediction, and terminal yard planning on the data shapes Hapag-Lloyd, Hamburg Süd, and HHLA operate. Models consume AIS, weather (DWD and ECMWF), bunker-price feeds, ULSFO and VLSFO availability, and ship-internal telemetry over CIP and OPC UA bridges. We respect IMO 2023 Carbon Intensity Indicator (CII), FuelEU Maritime (Regulation (EU) 2023/1805), and EU ETS extension to shipping, producing the carbon-accounting outputs the Bundesamt für Seeschifffahrt und Hydrographie (BSH) and the European Maritime Safety Agency (EMSA) expect. Reefer pipelines ship with §203 StGB-clean handling for any pharmaceutical cold-chain cargo and a documented rollback plan vessel masters can execute offline.

We build size-prediction and recommendation models on returns-history, browsing, basket, and explicit size-feedback signals, plus body-pose virtual try-on with MediaPipe, OpenPose, and on-device WebAssembly inference. Hamburg fashion AI must respect Article 22 GDPR (automated decisions producing legal or similarly significant effects need a human-review pathway for size-driven exclusion or price personalisation), DSGVO Article 9 when body measurements identify a person, and the new TDDDG (Telekommunikation-Digitale-Dienste-Datenschutz-Gesetz) rules on tracking consent. Models train and infer inside eu-central-1, with returns-reduction A/B testing instrumented for the existing Snowplow, Adobe, or GA4 stacks Otto and About You typically run. Live deployments include drift monitoring and a fairness audit on protected attributes.

Hamburg insurers (Hanse-Merkur, Signal Iduna, HDI, and Hamburg Allianz subsidiaries) and banks (Hamburg Commercial Bank, M.M.Warburg, Donner & Reuschel) are BaFin-regulated. VAIT (Versicherungsaufsichtliche Anforderungen an die IT) and BAIT (Bankaufsichtliche Anforderungen an die IT) demand documented model governance, ICT risk management, and outsourcing controls. DORA (Regulation (EU) 2022/2554), applicable since 17 January 2025, layers ICT third-party risk, incident reporting within four hours of classification, and digital operational resilience testing. Our regulated builds run a VAIT/BAIT and DORA gap analysis at discovery, document the AI system in the BaFin meldepflichtige Auslagerung register where threshold triggers apply, and ship with the Annex IV EU AI Act documentation, model risk policy, and rollback runbook your internal audit needs.

Yes. We build computer vision for acne, eczema, hyperpigmentation, melasma, and routine skin-condition assessment using EfficientNet, ConvNeXt, and Segment Anything 2 backbones fine-tuned on consented Fitzpatrick-balanced datasets. Where the product is positioned as wellness or cosmetic guidance, it stays outside MDR. Where it claims diagnostic value, intended use crosses into Regulation (EU) 2017/745 MDR Class IIa, and we run the build through IEC 62304 software lifecycle, ISO 13485 quality management, ISO 14971 risk management, and a notified-body conformity assessment (typically TÜV Süd or BSI). All training and inference stay inside eu-central-1 or Azure Germany West Central under Article 9 DSGVO consent, with on-device inference where the mobile-app form factor allows.

We default to AWS eu-central-1 (Frankfurt), Azure Germany West Central (Frankfurt), and GCP europe-west3 (Frankfurt) for storage, training, and inference. For LLMs that must stay in Germany or the EU we use Aleph Alpha Luminous on Heidelberg infrastructure, Cohere's EU endpoints, Azure OpenAI Germany West Central, or self-hosted Llama 3.3 and Mistral on H100 clusters in Frankfurt or Hamburg. International transfer to US-hosted APIs is acceptable only when a Transfer Impact Assessment under Schrems II signs off and a Data Privacy Framework recipient is selected. We document residency in the model card, the Verarbeitungsverzeichnis, and the Auftragsverarbeitungsvertrag, so the Hamburgische Beauftragte für Datenschutz und Informationsfreiheit (HmbBfDI) and BaFin reviewers have a clear answer.

Yes. We have scoped predictive-maintenance pipelines on engine performance, vibration spectra, oil-debris monitoring, and APU telemetry for aviation MRO operators. Models ingest ACMS and ACARS feeds, line-maintenance records, and OEM service-bulletin context, and they output remaining-useful-life and removal-recommendation scores under EASA Part-145 and Part-CAMO documentation requirements. Each model ships with a flight-safety hazard analysis aligned with ARP4761A and a configuration-management trail an EASA auditor can follow. Outputs feed AMOS, TRAX, or Ramco MRO stacks. We do not certify anything as airworthy software, but we build the AI layer to integrate cleanly with the human-in-the-loop authority a release-to-service holds.

A typical Hamburg production AI build (route optimisation, demand forecast, dermatology vision baseline, claims triage, MRO predictive maintenance) takes fourteen to twenty-four weeks from kickoff to live, assuming clean historical data and DSGVO and EU AI Act documentation are in scope. Week 1 to 4 covers data audit, regulatory classification, and baseline. Week 5 to 12 is modelling, iteration, and challenger testing. Week 13 to 18 is MLOps on eu-central-1, monitoring, shadow-mode deployment, and works council and Datenschutzbeauftragte review. Week 19 onward is gradual rollout. BaFin VAIT or MDR Class IIa pathways add eight to twelve weeks for conformity assessment and notified-body interaction. Pre-data or labelling-heavy projects add six to eight weeks.