L1Live
Workspace records
Customers, invoices and payments stored per company workspace. Access is enforced by row-level database policies tied to workspace membership, not by the interface alone.
Architecture & Trust
How DueWise turns records into intelligence.
Four layers, each with a clear boundary. The deterministic engine is the source of truth, a workspace-trained model scores late-payment risk, and AI explains both from your own workspace data.
L2Live
Deterministic engine
Settlement, ageing, customer payment behaviour, collection likelihood, priorities and 7/30/90-day forecasts. Same records and date always give the same output.
L3Live
Late-payment risk model
Logistic regression trained on each workspace’s own settled invoices once at least 10 exist. Inputs: customer late rate, history depth, invoice amount and days overdue. Output: late-payment probability (0–100%) and an A–D grade per open invoice.
L4Live
AI analysis layer
A language model reads authenticated workspace aggregates — open balances, ageing, forecast and customer behaviour — and returns structured review guidance. It explains engine outputs and never writes to source records.
Model governance
Prediction is earned by data, not assumed.
The late-payment risk model (logistic regression, trained per workspace) follows these implemented rules. It predicts whether an invoice will be paid after its due date. Cash-flow forecasts always use the separate published rule-based calculation.
- Training
- Trains on the workspace’s fully settled invoices; each is labelled late or on time by comparing its final payment date with its due date.
- Activation
- Training and scoring use one requirement: at least 10 fully settled invoices. There is no separate deployment gate.
- Fallback
- Below 10, no risk scores are shown and the deterministic forecast continues unchanged.
- Evaluation
- Leave-one-out cross-validation: every settled invoice is scored by a model trained without it. Held-out accuracy, precision and recall for late payment (0.5 threshold) are reported with each model.
- Calibration
- Held-out Brier score is reported against a base-rate baseline, with predicted-versus-observed late rates in four probability bands. Metrics are recomputed whenever the workspace model retrains.
Data controls
Isolation
Every finance table is scoped to a workspace ID and protected by database policies.
Authentication
Authenticated sessions are required for every workspace page and server function.
Traceability
Forecast lines export with balance, expected date, probability and the rule applied.
AI boundary
AI output is labelled, read-only and generated only from the signed-in workspace’s data.
Portability
Customers, invoices and payments export to CSV at any time.