Explore the engineering
How does a forecast earn trust? Explore the JobAI workflow.
Interactive architecture · research in progress
Select a stage to see its input, decision and output. This is an architecture walkthrough; no model runs here.
Data
- Input
- Official quarterly vacancy figures from Statistics Finland.
- Engineering decision
- Keep numeric data separate from the language model. Record the query and response metadata so an experiment can be traced.
- Output / requirement
- A traceable data snapshot, rather than numbers generated from a prompt.
Forecast
- Input
- Historical vacancy observations, ordered by time.
- Engineering decision
- Establish seasonal naive, last-value, regression and Prophet baselines before comparing the fine-tuned model.
- Output / requirement
- Candidate forecasts for one, two and four quarters ahead.
Evaluate
- Input
- Baseline and model predictions for the same forecast horizons.
- Engineering decision
- Use rolling-origin evaluation. Later observations must not leak into the earlier training window.
- Output / requirement
- The model must beat the baselines. No winning result is claimed while evaluation is in progress.
Try a forecasting baseline
Move the training cutoff. Compare what each method predicts for the next quarter.
Filled: training · outlined: held out · teal: forecast
- Next forecast
- Held-out value
- Absolute error
Illustrative quarterly values, not labour-market data or JobAI results. Only observations before the cutoff enter the forecast.
Explain
- Input
- Forecast results and approved official publications.
- Engineering decision
- Retrieve source material for the narrative. Keep the explanation grounded in that evidence, separate from numeric forecasting.
- Output / requirement
- A planned explanation layer with traceable sources. It cannot establish forecast accuracy by itself.