Biography

I am an AI researcher and engineer specializing in Computer Vision, AI Safety, and Document Intelligence. My research focuses on bridging the gap between theoretical AI safety concepts (Adversarial ML) and their robust application in high-compliance environments.

Currently, I operate as an independent Lead AI Engineer, designing specialized OCR and vision architectures for complex industrial domains (Engineering, Legal, Finance). Previously, I served as a Project Engineer for System & Integration Testing at MAN Truck & Bus SE, where I led the HiL Security team and focused on automotive cybersecurity. I hold an M.Sc. in Cyber Security from HDBW Munich.

News & Updates

  • Dec 2025 Released German-OCR-Engine (Tech Preview), a native C++ research artifact for local document analysis.
  • Dec 2025 Published German-OCR-Turbo (Qwen3-VL-2B fine-tune) - 3x faster than DeepSeek-OCR.
  • Jan 2025 Identified CVE-2025-0990 (CSRF Vulnerability).
  • Nov 2024 Published SecIDS-CNN model on Hugging Face (97.72% accuracy).
  • Oct 2024 Completed Master's Thesis on Trojaned AI Models.
  • May 2024 Started Team Lead position at MAN Truck & Bus (HiL Security).

Research Artifacts: Document Intelligence

Development of specialized Vision-Language Models (VLMs) and inference engines.

German-OCR Engine (Framework)

Project Page

A proprietary high-performance framework for local document analysis. Investigates the efficacy of quantized VLMs for extracting structured data from complex German documents (schematics, contracts). Features AES-256 model encryption and RAM-only inference.

C++Qt6Computer VisionQuantization

German-OCR Models (Open Research)

Hugging Face GitHub

A family of Qwen-based VLMs fine-tuned on synthetic German document datasets. Optimized for structured output (JSON/Markdown) and edge deployment via Ollama/Llama.cpp.

PythonPyTorchSynthetic DataLoRA

AI Safety & Security Research

SecIDS-CNN & Automotive Security

Model Code

Development of CNN and LSTM-based Intrusion Detection Systems (IDS) for automotive Controller Area Networks (CAN). Detects Fuzzy and DoS attacks with >97% accuracy.

TensorFlowCAN BusEmbedded Security

Security Vulnerability Disclosures (CVEs)

Official NIST Records

Responsible disclosure of critical security vulnerabilities in widely used software systems.

CVE-2021-4455 Remote Code Execution (RCE)
CVSS 9.8
CVE-2021-24997 Information Disclosure
CVSS 6.5
CVE-2025-0990 Cross-Site Request Forgery (CSRF)
CVSS 4.3

Open Source Engineering

AI-Driven SIEM

Real-time threat detection system using CNNs and NLP with Groq integration.

Autonomous XGO Robot

Navigation system for quadruped robots using computer vision and object detection.

Acoustic Echo Cancellation

Adaptive filter algorithms (LMS, RLS) for audio signal processing.

Vulna v4.0

AI-Powered Penetration Testing Framework.

Selected Publications

  • Manipulation of Neural Language Models: Security Risks and Defense Strategies for Trojaned AI Models
    K. Hardani • Master's Thesis, HDBW Munich, 2024
  • Verbesserung von Belohnungsmodellen durch Adversariale Trigger-Generierung
    K. Hardani et al. • Technical Report, 2024

Professional Experience

2017 – Present

Independent AI Research Engineer

Applied Research (Self-Employed)

Designing OCR & Computer Vision architectures for specialized industries.

2023 – 2025

Project Engineer - HiL & AI Security

MAN Truck & Bus SE (via Vdynamics)

HiL Security team, developed Reinforcement Learning Firewalls.

2021 – 2023

Senior Full-Stack & AI Developer

Kapital Medien GmbH

Developed AI-powered financial modeling engines.