German-OCR 4 · live Sep 2026

Applied AI Researcher.
Engineer. Builder.

I build the German AI stack — compact, local-first vision-language models, agentic systems, and LLM infrastructure. Research-grade, production-shipped, open-source-first.

🔬 AI Researcher 💼 AI Engineer
6
Published CVEs
(4× Critical)
100+
Public repos
on GitHub
ICLR'26
TurboQuant integration
for llama.cpp
0 %
Hallucination · 200+
German invoices tested
Premium · live since September 2026

German-OCR 4

Document AI for German invoices, plans and scans — text, structure and fields, each value with its source location in the original, handed off to SAP, RIB or your DMS. Unreadable fields stay empty: no guessed values. Available exclusively as a premium service through german-ocr.de.

Output
JSON · XML · PDF
Availability
Premium only
Guessed values
None

Open-source predecessor: German-OCR 3.1 (Apache 2.0) · Hugging Face · Ollama · GitHub

Security Research

6 published CVEs · 10 advisories in triage

Responsible-disclosure security research via NVD, GitHub Security Lab, Patchstack and vendor programs. Targets include enterprise-grade systems at Siemens, Tutanota, Mailcow, Shopware, Pimcore, Einride and NousResearch.

CVE Severity Type Target Year
CVE-2026-404929.8 CriticalNative RCE vectorSAIL image library2026
CVE-2026-404939.8 CriticalNative RCE vectorSAIL image library2026
CVE-2026-404949.8 CriticalNative RCE vectorSAIL image library2026
CVE-2021-44559.8 CriticalArbitrary File Upload · RCEWordPress · Smart Product Review2021
CVE-2021-249976.5 MediumAuth / Info DisclosureWordPress · WP Guppy2021
CVE-2025-09904.3 MediumCSRFWordPress · I Am Gloria2025
In Triage · Pending Disclosure
  • Siemens — KAS · linux-entra-sso (2 findings)
  • NousResearch — hermes-agent (2 findings · incl. Critical)
  • Tutanota — end-to-end encrypted mail
  • Mailcow — dockerized mail server
  • Shopware — commerce platform (2 findings)
  • Pimcore — admin-ui-classic-bundle
  • Einride — iam-go identity framework
Certifications & Disclosure
  • OSCP · Offensive Security Certified Professional
  • CEH · Certified Ethical Hacker
  • ISTQB · Certified Tester Foundation
  • ISO 21434 · Automotive Cybersecurity (Vector)

Responsible disclosure only via NVD, GitHub Security Lab, Patchstack. No 0-day sales, no PoC leakage.

Research

Academic work · Publications

Research report · Version 1.0 · 7 November 2026

Weight-Space Programming with Claude Mythos: Weight Edits and Implanted Circuits in a 0.8B Language Model, with Transfer to an Engineering Surrogate

Keyvan Hardani · Experiments 3–7 November 2026

Claude Mythos 5.1 edits the weights of Qwen3.5-0.8B directly, without training.

  • Edit 1 (10 passes): rank-3 projection in three attention matrices. MATH-500 from 7.3 % to 32.5 %.
  • Edit 2 (8 further passes): four silent MLP units repurposed into a loop detector with early stop. MATH-500 58.7 %, truncated answers from 66 % to 17 %.
  • Reference: GPT-6 Astra reached 48.1 % with 50 passes in the MIT study (Takehi, Teoh, Zhang, Khattab, MIT CSAIL, 2 November 2026).
  • Transfer: thermal surrogate model of a power module. With 16 part-load simulations, 0.17 K error at full load, versus 2.3 to 4.5 K for fine-tuning.
Applied Research

AI for Engineering Systems

Focus: Intelligent document processing, agent-based systems for civil engineering and industrial applications.

MSc · 2025

Cybersecurity (HDBW Munich)

Master's Thesis

"Manipulation of Neural Language Models — Security Risks and Defense Strategies for Trojaned AI Models"

Publications
Stack

What I use to build this

Languages
Python · C/C++ · Go · TypeScript · SQL · Rust
ML / AI
PyTorch · Transformers · PEFT/LoRA · vLLM · llama.cpp · Ollama · ONNX
Agentic / LLM
LangChain · RAG · Vector DBs (Weaviate, FAISS, Pinecone) · Fine-tuning
Infra
Docker · CUDA · FastAPI · GPU orchestration · CI/CD · MLOps
Work with me

Advisory · Consulting · Research collaboration

Selectively taking on engagements where I can contribute deep technical work — agentic AI, document intelligence, LLM infrastructure, on-premise / DSGVO-compliant ML stacks, AI security.

Short advisory

Architecture reviews, model-selection, fine-tuning strategy, GPU sizing.

Research collaboration

Joint work on VLMs, document AI, agentic systems, AI safety.

Start a conversation