Makesense MLOps Pipeline

End-to-end multimodal MLOps pipeline for entity extraction, relationship discovery, and knowledge graph construction.

Makesense MLOps Pipeline presents a systematic study of transformer-based architectures for abstractive dialogue summarization, with emphasis on computational efficiency and deployment readiness.

Research Contributions

  • Systematic Architecture Comparison — Comprehensive evaluation of encoder-decoder variants (BART, T5, Pegasus) on dialogue summarization with matched compute budgets.
  • Efficiency-Quality Tradeoffs — Quantitative analysis of the Pareto frontier between ROUGE scores and inference latency.
  • Training Strategy Ablations — Investigation of warmup scheduling, gradient accumulation, and mixed-precision training.
  • Production-Ready Pipeline — End-to-end MLOps implementation with experiment tracking, model versioning, and reproducibility guarantees.

Links: GitHub