Built a data orchestration layer for a Jitsi meeting system that turns raw meeting transcripts into structured inputs for topic segmentation and recap generation.
Overview
The system ingests meeting transcripts, persists workflow state in PostgreSQL, and prepares model-ready data for segmentation and recap workflows. It also captures human corrections and edited recaps so feedback can be reused for model improvement.
Key Features
- Transcript ingestion and normalization for downstream model workflows
- PostgreSQL-backed workflow state tracking
- Transformation of raw meeting data into model-ready inputs
- Human-in-the-loop feedback persistence for user corrections and edited recaps
- Reusable feedback data for retraining and evaluation
- MLOps monitoring and quality gates for model degradation detection
Technical Implementation
The project uses Python for orchestration and data transformation, PostgreSQL for durable workflow state, Docker for reproducible deployment, and MLflow for experiment and artifact tracking. Grafana and Alertmanager provide monitoring coverage for data quality and model health signals.
Impact
The pipeline creates a practical foundation for improving meeting intelligence workflows over time through structured feedback, measurable quality checks, and repeatable model operations.