Overview
This system summarizes chat-like texts using a fine-tuned PEGASUS model on the SAMSum dataset. It demonstrates a complete machine learning lifecycle: data curation, training, version control, CI/CD deployment, and inference APIs.
My Role & Contributions
- Data & Schema: Chose SAMSum, defined I/O schema.
- Modeling: Fine-tuned PEGASUS, evaluated via ROUGE.
- Pipelines: Modular ML pipelines with config-driven architecture.
- Deployment: FastAPI app, Dockerized, CI/CD with GitHub Actions + AWS EC2.
- Monitoring: Logging, inference endpoint with live request handling.
Tech Stack
Python
FastAPI
Docker
AWS EC2
GitHub Actions
Transformers
Tokenizers
Matplotlib
Implementation Details
- Modular Structure: Separated components into
components/
,pipeline/
,config/
. - Config-Driven: Reproducible via
config.yaml
+params.yaml
. - Inference Server: FastAPI
/predict
endpoint with JSON input/output. - Docker & CI/CD: GitHub Actions, AWS ECR, auto-deployed to EC2.


Results & Impact
- Achieved ROUGE-1 F1 ≈ 45 on SAMSum summaries.
- Reduced verbose conversations by 60–70% while retaining meaning.
- CI/CD reduced deployment from hours to minutes.
- Scalable architecture ready for containerized environments.