KIVEDU

My Role: Technical Project Lead

Technologies

Project Overview

KIVEDU is a research project funded by the Ministry of Economy, Labor, and Tourism of Baden-Württemberg, developed in collaboration with XPACE GmbH, Law & Audrey UG, and the University of Pforzheim. The project addresses a growing issue in e-commerce: identifying and enforcing consumer rights violations. Using AI-powered tools, KIVEDU automates tasks like detecting legal infractions, securely archiving evidence, and presenting actionable insights for consumer protection agencies.

As the technical project lead, I was responsible for designing the system architecture, leading a team of four developers, and ensuring the platform met both technical and regulatory requirements. One key aspect of the project was building a solution flexible enough to run in the cloud during development but fully on-premises for production, including hosting self-managed LLMs.

My Role

• System Design: Designed the platform architecture to be scalable, event-driven, and modular, supporting both cloud and fully on-premises deployments. This included configuring infrastructure for hosting a self-managed open-weight LLM during testing.
• Team Leadership: Led a team of four developers, overseeing implementation, assigning tasks, and ensuring deadlines were met.
• Web-Based Platform Development: Built the platform using modern web technologies like Nuxt.js and Supabase, with a focus on user experience and efficient workflows.
• AI-Powered Analysis: Integrated LLMs to automatically identify violations of cease-and-desist orders in German product descriptions, achieving high accuracy in legal infraction detection.
• Semantic Web Scraping: Developed tools for crawling and scraping e-commerce platforms to extract and analyze product data.
• Immutable Archival: Implemented a secure archival system for storing evidence, including full-page screenshots and static exports, ensuring tamper-proof data storage for legal purposes.
• Event-Driven Workflow: Used event-driven processing to enable scalable and asynchronous operations, utilizing Azure services during development and transitioning to on-premises infrastructure for production.
• LLM Research: Conducted research into optimizing LLMs for consumer rights enforcement, leading to three published papers presented at international conferences.

Technical Project Description

KIVEDU is built as a distributed microservices system with asynchronous event-driven processing. It consists of multiple components:

• Backend: Built with Supabase (PostgreSQL), with triggers that create jobs for processing new entities.
• Worker Infrastructure: Uses Graphile Worker to handle processing queues and execute various asynchronous tasks.
• Crawling & Scraping: Uses Crawlee and Playwright to extract data from e-commerce stores.
• AI Models:
  • Text Extraction & Information Retrieval: OCR and structured data extraction from uploaded PDFs of cease-and-desist declarations.
  • Violation Detection: Uses LLMs for classifying whether an e-commerce website violates consumer rights.
  • Title Generation: AI-generated semantic document titles for user-friendly overviews.
• Frontend: Built with Nuxt.js and Vue.js, styled with Tailwind CSS and DaisyUI, providing a seamless UI for users.
• Deployment & Infrastructure:
  • Development in the cloud using Azure.
  • Production on-premises for privacy-sensitive environments, including self-hosted LLMs using vLLM.

Challenges

• Scalability: The system needed to support large-scale web crawling and scraping without breaking under load.
• AI Model Accuracy: Optimizing violation detection models required extensive evaluation, prompt tuning, and fine-tuning on domain-specific datasets.
• Data Privacy & Security: The system was designed for full on-prem deployment due to regulatory constraints, requiring specialized infrastructure setups.
• Archival Integrity: Ensuring that archived evidence was immutable and could be presented in legal contexts without risk of tampering.
• User Experience: Designing a UI that made complex legal workflows accessible to non-technical users.

Achievements

• High-Accuracy Violation Detection: Developed AI models capable of detecting violations with ~90% accuracy.
• Scalable Event-Driven System: Designed a backend architecture that scales seamlessly as workloads increase.
• On-Premises Deployment Success: Successfully transitioned from cloud-based development to fully on-premises production infrastructure.
• Research Publications: Published three research papers on AI for consumer rights enforcement, presented at international conferences.
• Stakeholder Engagement: Presented the system to consumer protection agencies and industry professionals, generating significant interest.

Current Status

The KIVEDU platform is currently in the final stages of testing, transitioning fully to on-prem infrastructure for production deployment. The project has been well received by stakeholders, and ongoing refinements are being made to further improve AI accuracy and user workflows.