🚀 Astronomical Source NER Pipeline

🧠 Project Overview

This project develops a custom Named Entity Recognition (NER) model that extracts references to celestial sources (e.g., quasars, black holes, gamma-ray bursts) from scientific abstracts and real-time alerts.

The model is trained using astronomy abstracts from NASA ADS and then deployed to identify source objects in the Astronomer's Telegram (ATel) RSS feed. It is designed with full MLOps capabilities, including experiment tracking, workflow orchestration, model deployment, monitoring, and reproducibility.

---

🎯 Problem Description

In astrophysics, the discovery and study of celestial sources are documented in scientific literature and real-time alert networks like ATel. However, identifying and indexing source names from this unstructured data is challenging due to inconsistent formatting, varied nomenclature, and volume.

This project addresses that by:

• Creating a labeled dataset of astronomical sources from NASA ADS abstracts
• Fine-tuning a spaCy NER model to recognize source names
• Applying the model to ATel RSS feed entries to extract entities
• Providing a Streamlit web app for real-time interaction

---

☁️ Cloud Architecture

We use AWS as the cloud platform and Terraform for Infrastructure as Code (IaC) to provision and manage resources.

Cloud Components

Component	Technology
Cloud Provider	AWS (S3, Lambda, CloudWatch)
IaC	Terraform
Data Storage	S3
Workflow Scheduler	Prefect Cloud (Free Tier)
Model Hosting	AWS Lambda
Monitoring	Grafana, Evidently, CloudWatch
CI/CD	GitHub Actions
Web App	Streamlit (locally or on ECS)

✅ This setup meets the 4-point cloud criterion: fully cloud-hosted with IaC.

---

🧪 Experiment Tracking and Model Registry

All experiments are tracked with MLflow, which logs:

• spaCy model versions
• Dataset versions
• Hyperparameters and metrics (precision, recall, F1)

A model registry tracks models in staging and production.

✅ This setup meets the 4-point experiment tracking criterion: tracking + registry used.

---

🔁 Workflow Orchestration

Workflows are built and scheduled with Prefect Cloud:

Flows:

1. build_dataset: Parse and label NASA abstracts
2. train_model: Train spaCy NER model
3. evaluate_model: Run evaluation and log metrics
4. inference_atel: Poll ATel RSS feed and extract sources
5. monitor: Run drift reports with Evidently

✅ Fully deployed with remote Prefect agent and schedules.

---

🚀 Model Deployment

• Batch inference is performed via AWS Lambda (low-cost and scalable).
• spaCy model is compiled to optimize Lambda performance.
• Outputs written to S3 and logged to CloudWatch.

Streamlit Demo

A Streamlit web app enables users to:

• Paste in text (e.g., ATel message)
• View extracted source names
• Can run locally or be deployed to AWS ECS

✅ Meets 4-point deployment criterion: containerized + cloud-executable.

---

📈 Monitoring

Monitoring is conducted via:

• Evidently for drift detection and label distribution checks
• Grafana dashboards via CloudWatch metrics (cold starts, errors, latency)

Conditional Flow

If metrics degrade (e.g., F1 drop, drift detected):

• Prefect triggers a retraining flow
• Slack/email alerts notify maintainers

✅ Meets 4-point monitoring criterion: alerts and conditional logic included.

---

🧹 Reproducibility

• Project uses uv for fast, modern dependency and environment management
• Code and data are versioned
• All flows are reproducible via Prefect
• Infrastructure managed by Terraform

✅ Meets 4-point reproducibility criterion.

---

✅ MLOps Best Practices

Practice	Status
Unit tests	✅ pytest in tests/unit/
Integration tests	✅ pytest in tests/integration/
Linter/Formatter	✅ black, ruff
Makefile	✅ Automation for test/lint/train
Pre-commit hooks	✅ pre-commit used
CI/CD	✅ GitHub Actions

✅ Full 7/7 points for MLOps best practices.

---

🗂️ Project Structure

astronomical-ner/
├── data/                      # Raw + processed ADS abstracts
├── notebooks/                 # EDA and dataset building
├── src/
│   ├── data_ingestion/        # Download from NASA ADS
│   ├── labeling/              # Create spaCy-compatible annotations
│   ├── training/              # Model training logic
│   ├── inference/             # ATel RSS + entity extraction
│   ├── monitoring/            # Evidently integration
│   ├── app/                   # Streamlit demo app
├── tests/
│   ├── unit/                  # Unit tests
│   ├── integration/           # End-to-end tests
├── deployment/
│   ├── lambda/                # Lightweight Lambda inference logic
│   ├── docker/                # Container setup for app and model
├── terraform/                # IaC: S3, Lambda, IAM, CloudWatch
├── .github/workflows/        # CI/CD pipelines
├── Makefile
├── requirements.txt
├── pyproject.toml
├── .pre-commit-config.yaml
├── README.md

# 🧪 How to Run (Dev Mode) — Using uv
## Install uv (if not installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

## Install dependencies
uv pip install -r requirements.txt

## Lint + format code
uv pip install black ruff
black .
ruff check . --fix

## Run all tests
uv pip install pytest
pytest

## Run specific test sets
pytest tests/unit/
pytest tests/integration/

## Launch Streamlit app
streamlit run src/app/app.py