AI-powered news intelligence and retrieval platform built with FastAPI, PostgreSQL, machine learning, semantic embeddings, and hybrid retrieval workflows.
SignalStack ingests RSS news feeds, normalizes and enriches article data, stores semantic embeddings locally, separates training labels from model predictions, and exposes operational APIs plus a lightweight frontend dashboard for search, ingestion, retrieval, and AI/ML demonstration workflows.
Video walkthrough: watch the SignalStack demo
Frontend dashboard:
Interactive API docs:
SignalStack was built as a portfolio project to demonstrate backend software engineering, database-backed search, AI/ML enrichment workflows, semantic retrieval, and operational dashboard design.
The project is intended to support applications for entry-level software engineering, backend engineering, AI software engineering, AI platform engineering, and database-heavy application development roles.
- Backend API development with FastAPI
- PostgreSQL-backed application design
- Database migrations with Alembic
- RSS ingestion and data normalization
- Full-text, semantic, and hybrid retrieval workflows
- Applied machine learning classification with scikit-learn
- Local sentence-transformer embedding generation
- Clear separation between ML training labels and inference predictions
- Lightweight frontend dashboard development
- Operational visibility for ingestion and AI workflows
SignalStack is currently portfolio-demo ready.
Implemented functionality includes RSS ingestion, article storage, API search/filtering, full-text search, semantic search, hybrid retrieval, ML classification, embedding persistence, frontend dashboard views, and operational RSS feed controls.
The roadmap below describes possible future enhancements rather than required functionality for the current demo.
- RSS feed ingestion pipeline
- Configurable feed source management
- Duplicate detection
- Article normalization and text cleaning
- RSS ingestion timeout handling
- Operational ingestion dashboard
- Feed synchronization between config and database
- Feed health/error visibility
- Live ingestion triggering from frontend
- PostgreSQL TSVECTOR full-text search
- Weighted ranking with
ts_rank_cd - Phrase search
- Keyword filtering
- Metadata filtering
- Semantic embedding search
- Related article retrieval
- Hybrid keyword + semantic ranking
- Pagination and sorting
- Local/offline sentence-transformer embeddings
- scikit-learn article classification
- TF-IDF vectorization
- Naive Bayes classification
- Weak-supervision training-label workflow
- Separation of training labels from model predictions
- ML confidence scoring
- Persisted vectorizers and model artifacts
- Live inference API
- Article similarity retrieval
- Embedding persistence in PostgreSQL
- Static frontend served by FastAPI
- Dashboard statistics cards
- Sticky operational navigation
- RSS feed operational modal
- Live ingestion controls
- RSS synchronization status messages
- Semantic search interface
- Hybrid search interface
- ML prediction testing
- Related article lookup
- Live dashboard refresh after ingestion
- RSS feed source synchronization
- Ingestion metrics and visibility
- Feed activation/deactivation
- Last ingest timestamps
- Error propagation to frontend
- Offline-capable AI architecture
- Local ML artifact loading
- Service-layer organization
- Dockerized PostgreSQL deployment
- Python 3.13
- FastAPI
- SQLAlchemy ORM
- Alembic
- PostgreSQL 16
- Docker
- scikit-learn
- sentence-transformers
- semantic embeddings
- cosine similarity retrieval
- TF-IDF vectorization
- Naive Bayes classification
- weak-supervision labeling
- HTML
- CSS
- JavaScript
SignalStack follows a modular service-layer architecture designed for AI-enhanced retrieval, operational ingestion workflows, and offline-capable ML pipelines.
RSS Feeds
→ Ingestion Pipeline
→ Article Cleaning & Enrichment
→ Weak-Supervision Training Labels
→ ML Classification Inference
→ Semantic Embedding Generation
→ PostgreSQL Storage
→ Search / Retrieval APIs
→ Frontend Dashboard
The RSS ingestion pipeline synchronizes configured feed sources, performs ingestion, enriches article content with ML classification and embeddings, and updates operational dashboard statistics.
SignalStack supports keyword search, semantic retrieval, and hybrid ranking by combining PostgreSQL full-text search with local embedding similarity workflows.
SignalStack separates weak-supervision training labels from model prediction outputs. Training labels are used to generate classifier artifacts, while inference uses those artifacts to produce ml_category and ml_confidence values.
Core architectural goals:
- offline-capable AI workflows
- reusable service-layer logic
- modular API endpoints
- explainable retrieval systems
- clear training vs inference separation
- incremental platform evolution
- lightweight deployment footprint
SignalStack includes a lightweight frontend dashboard served directly by FastAPI.
Frontend URL:
http://127.0.0.1:8000/demo
The dashboard supports:
- viewing live platform statistics
- triggering RSS ingestion
- viewing RSS feed synchronization status
- keyword article search
- semantic embedding search
- hybrid retrieval search
- ML prediction testing
- RSS feed operational visibility
- related article lookup
The frontend intentionally uses plain HTML, CSS, and JavaScript to keep the project lightweight and focused on backend engineering, AI retrieval systems, and operational workflows.
GET /api/v1/articles
GET /api/v1/articles?search=ai
GET /api/v1/articles?phrase_search=artificial intelligence
GET /api/v1/articles?min_ml_confidence=0.70
GET /api/v1/articles/semantic-search?query=artificial%20intelligence%20semiconductor%20chips
GET /api/v1/articles/hybrid-search?query=cybersecurity%20ransomware%20hospital
GET /api/v1/articles/{article_id}/similar
POST /api/v1/ml/predict
GET /api/v1/ml/health
GET /api/v1/rss-feeds
POST /api/v1/rss-feeds/sync
POST /api/v1/rss-feeds/ingest
GET /api/v1/dashboard/stats
- Total articles
- Active RSS feeds
- ML-classified articles
- Embedded articles
- Last article ingest timestamp
- Last feed ingest timestamp
app/
├── api/
├── config/
├── crud/
├── db/
├── ingestion/
├── ml/
├── models/
├── schemas/
├── services/
├── static/
└── utils/
models/
└── ml/
├── article_classifier.joblib
└── article_vectorizer.joblib
requirements/
└── requirements.txt
docs/
├── diagrams/
├── screenshots/
└── srs/
git clone <your-repository-url>
cd signalstackpython -m venv .venv
source .venv/bin/activatepip install -r requirements/requirements.txtcp .env.example .envAPP_NAME="SignalStack API" ENV=dev DATABASE_URL=postgresql+psycopg://signalstack_user:signalstack_password@localhost:5432/signalstack
docker compose up -dpython -m alembic upgrade headpython -m app.ingestion.fetch_rsspython -m app.ml.backfill_training_labelspython -m app.ml.train_article_classifierpython -m app.ml.backfill_ml_confidenceFor normal demo/runtime use:
python -m uvicorn app.main:appFor active development with automatic reload:
python -m uvicorn app.main:app --reloadNote: --reload is useful during active code editing, but it can increase CPU usage on some Linux laptop environments because Uvicorn watches the filesystem for changes.
http://127.0.0.1:8000/demo
http://127.0.0.1:8000/docs
SignalStack separates machine learning training workflows from inference workflows.
Training means generating model artifacts from labeled article data.
training_label + article text
→ TF-IDF vectorizer
→ Naive Bayes classifier
→ persisted .joblib artifacts
Inference means loading trained artifacts and using them to make predictions on article text.
article text
→ vectorizer artifact
→ classifier artifact
→ ml_category + ml_confidence
This distinction prevents model predictions from being reused as training labels and helps avoid self-reinforcing feedback loops.
SignalStack stores independent training labels separately from model prediction outputs.
training_label
represents a weak-supervision label used for training.
training_label_source
describes how that label was generated.
Current training labels are generated using a deterministic rule-based weak-supervision classifier.
This supports a cleaner ML lifecycle:
rule-based weak supervision
→ training_label
→ classifier training
→ model artifacts
→ inference
→ ml_category + ml_confidence
ml_category
represents the classifier prediction produced during inference.
ml_confidence
represents the model confidence score associated with that prediction.
These values are model outputs, not ground-truth labels.
SignalStack persists lightweight scikit-learn artifacts under:
models/ml/
Current artifacts:
models/ml/article_classifier.joblib
models/ml/article_vectorizer.joblib
These artifacts are committed to the repository because they are small enough for the current project and support reproducible local inference.
Larger model assets, such as sentence-transformer model files, are intentionally managed outside the repository when appropriate.
SignalStack uses locally hosted sentence-transformer models for semantic embedding generation.
Embeddings are currently stored in PostgreSQL JSON columns for portability and development simplicity.
Future upgrade path:
- pgvector integration
- approximate nearest-neighbor retrieval
- distributed embedding pipelines
- local model registry organization
- GPU-accelerated inference support
Hybrid search combines:
- PostgreSQL full-text ranking
- semantic embedding similarity
- weighted scoring strategies
to improve search relevance and contextual retrieval quality.
- lightweight deployment
- explainable retrieval systems
- offline-capable AI workflows
- modular architecture
- clear training vs inference separation
- operational visibility
- local reproducibility
- incremental engineering evolution
- recruiter/demo friendliness
- pgvector integration
- ANN retrieval optimization
- async ingestion workers
- streaming ingestion pipelines
- article summarization
- entity extraction
- topic modeling
- improved weak-supervision label coverage
- human-reviewed training labels
- model evaluation reports
- model registry workflows
- automated retraining workflows
- authentication/admin controls
- deployment automation
- ingestion scheduling
- frontend visual analytics
Additional project documentation:
docs/diagrams/signalstack_overview.pumldocs/diagrams/rss_ingestion_flow.pumldocs/diagrams/ai_retrieval_flow.pumldocs/diagrams/ml_training_inference_flow.pumldocs/screenshots/docs/srs/signalstack_srs_outline.md
MIT License
Portfolio and educational project.
RSS feed content, headlines, excerpts, and associated media remain the property of their respective publishers and content owners.
SignalStack stores and processes publicly available RSS metadata and article excerpts for educational, research, portfolio, and software engineering demonstration purposes.
This project is intended to demonstrate backend engineering, AI retrieval workflows, semantic search, machine learning enrichment pipelines, weak-supervision training workflows, and operational dashboard architecture.










