Skip to content
Back to overview

AI-Powered Support Automation / RAG System

Logify for BCT – AI Log File Analysis with RAG for IT Support

AI log file analysis for BCT: Logify searches logs and around 500 MB of technical docs via RAG and suggests fixes. Estimated potential: more than 1,000 support hours per year. Running on-premise at the client since April 2026.

Logify for BCT – AI Log File Analysis with RAG for IT Support

Starting point

In BCT’s IT support, staff analyzed large, unstructured log files by hand to find the source of errors. On top of that came searching the technical documentation: just finding the right document among hundreds takes minutes every time.

Individual support cases can drag on accordingly, in extreme cases for weeks. The working hypothesis of the project: around 80% of cases have occurred before in a similar form. So the solution often already exists, it just cannot be found quickly.

Solution

Logify automatically analyzes log files, detects errors and anomalies, and generates prioritized solution suggestions with context. Via RAG, the system draws on around 500 MB of technical documentation in PDF format and brings the relevant passages into its suggestion.

The first version was built in 48 hours at the Black Forest Innovation Hackathon in November 2024 and won an award there. Then came the longer part, the road to production:

  • Model outputs are converted into a fixed, validated format (Pydantic) instead of being passed on as free text
  • Results land in a dedicated interface for the support team
  • The application and indexing run on-premise on the client’s infrastructure; the language model is connected via the OpenAI API and could be swapped for a local model

Why this second step almost always takes longer than the prototype is explained in our post AI implementation: from pilot to production (German).

Technology stack

  • Python (backend / analysis engine)
  • OpenAI LLMs (error and anomaly detection)
  • ChromaDB (RAG indexing and retrieval)
  • Pydantic (data validation / structured output)
  • TypeScript, ReactJS (frontend / support interface)
  • On-premise operation of application and index at the client

Result & Impact

  • Award at the Black Forest Innovation Hackathon (10/2025), prototype in 48 hours
  • On-premise deployment completed in April 2026, next step is internal use by the support team
  • Around $1 per 100 queries in model costs
  • Potential of more than 1,000 hours per year (estimate): at a conservative 30 minutes saved per support case. Robust measurements from production are still pending.
  • Follow-up project (HubSpot Support Copilot) in concept phase