How frozono9/ecityclic-recommender works
eCityclic is a Streamlit application that recommends administrative procedures from a user’s recent procedure history. A user selects an available procedure; the frontend maintains a short history and sends its sequence IDs to the backend. The backend loads the trained Transformer, pads the history, predicts likely next procedures, filters to active procedures, and resolves suggestions to titles for display. The README also describes notebook-based preprocessing and training. The referenced CSV is absent from the supplied tree, so it is shown as an external data dependency; notebook details are based on the README, not inspected source.
Main entry points: backend.py Recommendation entry · frontend.py Selection and display · modelo_transformer_tramites.h5 Trained Transformer · model.ipynb Preprocessing and training
Components 12
Streamlit interface
- Selection and display
streamlit/frontend.py - Recent history
streamlit/frontend.py
Recommendation runtime
- Recommendation entry
streamlit/backend.py - Model loader
streamlit/backend.py - Trained Transformer
model/modelo_transformer_tramites.h5 - Sequence padding
streamlit/backend.py - Transformer prediction
streamlit/backend.py - Active ranking
streamlit/backend.py - Title resolution
streamlit/backend.py
Model training
- Preprocessing and training
model/model.ipynb
Other components
- Procedure user
- Procedure catalog CSV
Connections 16
- Procedure user → Selection and display: selects procedure
- Selection and display → Procedure catalog CSV: reads titles and IDs
- Selection and display → Recent history: updates history
- Selection and display → Recommendation entry: sends sequence IDs
- Recommendation entry → Model loader: loads model
- Model loader → Trained Transformer: loads weights
- Recommendation entry → Sequence padding: pads sequence
- Recommendation entry → Transformer prediction: predicts sequence
- Transformer prediction → Trained Transformer: runs inference
- Recommendation entry → Active ranking: filters and ranks
- Active ranking → Procedure catalog CSV: reads active procedures
- Recommendation entry → Title resolution: resolves suggestions
- Title resolution → Procedure catalog CSV: looks up titles
- Recommendation entry → Selection and display: returns suggestions
- Selection and display → Procedure user: displays suggestions
- Preprocessing and training → Trained Transformer: trains and saves