Urja AI

Packed the full snapshot from the finished URJA-AI project.

Includes complete stack:

• Flask app with full API surface
• trained model artifacts
• extended + serving datasets
• training script + notebook
• full dashboard template and docs

Main goal here was not adding new features, just freezing the exact finished state so the whole progression has a clean end point.

Production-oriented checkpoint.

Backend and retraining script are now both in decent shape:

• dynamic latest model folder support
• full endpoint set (forecast, historical, metrics, annual, status)
• holdout + sequential CV in retraining
• cleaner dashboard flow for quick checking

This stage is where I’d normally hand over for testing before documentation and deployment files.

Created standalone retraining script so model updates are reproducible.

This first version trains and evaluates on the same set (not ideal), but it establishes:

• feature engineering pipeline
• artifact saving flow (model/scalers/config/metrics)
• version-based model output folder

Next step is proper holdout and sequential CV because current metrics are too optimistic.

Expanded backend API surface to something close to final product:

• /api/metrics
• /api/annual
• richer /api/status
• dynamic latest model folder selection

This made it easier to expose model quality and fiscal trend in UI.

The model folder auto-detect logic was a bit tricky because I wanted numeric sorting (nepal_2025 > nepal_2024) not simple string sort.

Added a proper browser dashboard so I can visually inspect both historical and forecast series.

This stage finally feels interactive:

• / renders HTML
• chart updates by calling API
• forecast horizon selectable (short/medium/long)

UI is still simple and not polished, but useful for validating model behavior quickly.

Switched forecast endpoint from naive baseline to actual ML inference.

Key things implemented:

• iterative step-by-step forecasting loop
• feature generation per target step
• scaler in/out transform around model prediction

Main headache was making sure lag features use updated predicted values in sequence. Once that was fixed, long-horizon forecast looked consistent.

Connected app to trained model artifacts for the first time.

This stage only proves loading works: model, scalers, config, and dataset.

I kept directory lookup hardcoded (models/nepal_2025) for now because I wanted a stable checkpoint before adding dynamic model version detection.

Before jumping to ML model loading, I added a naive forecast baseline (rolling mean of recent values).

This gives me a working /api/forecast and helps test the chart flow without model artifacts.

Forecast quality is obviously weak for seasonal spikes, but this stage was about plumbing, not accuracy.

Added dataset loading and basic status endpoint.

Now app can confirm if CSV is present and show row count/data range.

Small issue: parsing looked wrong at first because I forgot parse_dates=['date']. Fixed that and sorting is stable now.

Added dataset loading and basic status endpoint.

Now app can confirm if CSV is present and show row count/data range.

Small issue: parsing looked wrong at first because I forgot parse_dates=['date']. Fixed that and sorting is stable now.

Started with the smallest Flask app possible to make sure environment and routing are fine.

Nothing fancy yet, just a health-like response at /.

I almost started with modeling directly, but I’ve burned effort before when basic server setup was broken, so I’m keeping this boring on purpose.