keyvan.ai

What It Is

Keyvan.ai is the project and consulting homepage of an Applied AI researcher/engineer based in Munich, Germany, positioned around building a “German AI stack.” It is not a typical SaaS platform, but rather an entry point for a set of open-source projects and professional services focused on German document intelligence, local-first vision-language models, LLM infrastructure, and AI security.

Core Capabilities

The most prominent product on the site is German-OCR-3 v0.2: a compact Vision-OCR model for German documents, available in 1.0GB and 2.7GB versions. It is based on Qwen3.5, licensed under Apache 2.0, and can be accessed via Ollama, Hugging Face, and GitHub. On 200+ anonymized German invoices, it claims a 100% valid JSON rate, 95% sender identification, and 0% hallucination. Other projects include German-Privacy-Shield, German-Layout, TurboQuant-ggml, and KVCache-Autotune, covering PII redaction, layout analysis, and inference optimization.

Pricing and Integrations

The website does not disclose standard pricing, a free tier, or enterprise plans. Its commercial model appears to lean more toward consulting, architecture reviews, model selection, fine-tuning strategy, GPU sizing, research collaboration, and building on-premise/DSGVO-compliant ML stacks. In terms of integration, it mentions Ollama, llama.cpp, Hugging Face, FastAPI, Transformers, vLLM, LangChain, RAG, and vector databases, but there is no clear documentation for an official API, SDK, SLA, or admin console.

Pros and Cons

Its strengths lie in its very focused direction: German-language documents, local deployment, GDPR/DSGVO compliance, and lightweight inference, making it well targeted at European enterprise document workflows. Its open-source-first approach and Apache 2.0 licensing lower the barrier for experimentation. The author also combines AI engineering experience with a background in security vulnerability disclosure. The limitations are that this is not a mature commercial product site: it lacks pricing, payment options, customer support details, deployment boundaries, and customer case studies. The quality metrics for German-OCR-3 mainly come from German invoice samples, so its generalization to other document types still needs to be tested in practice.

Who It’s For and Access from China

It is better suited to companies and research teams with engineering capabilities that want to self-host German OCR/document AI, or that need LLM infrastructure and AI security consulting. It is less suitable for general users looking for an out-of-the-box, multilingual product with a full admin dashboard and commercial support. Access from China, payment methods, and Chinese-language support are not described in the main content, so china_access can only be considered unknown. If Chinese support or stable domestic alternatives are required, options to evaluate include PaddleOCR, Tesseract, Azure Document Intelligence, Google Document AI, AWS Textract, ABBYY, or Unstructured.