One engine under every decision.
Luminnovate turns streams of events about entities (customers, accounts, policies, members) into decisions with known outcomes. Most products ship a fixed library of attributes; Luminnovate builds the signals a decision needs, proves each one on the client's own outcomes, and keeps only those that earn their place.
Interpret the data
Each event is read for what it means for the entity: revenue, borrowing, owner money or a transfer. Identities are resolved across sources, histories are reconciled from overlapping extracts, and every figure is computed as of its date.
Illustrative example. Harbour Joinery is a synthetic business.
Create and test signals
Signals are proposed from hypotheses about how entities behave: rhythm, concentration, buffers, and changes against the entity's own normal. Each is measured over the window its cadence supports, then tested before any use: lift, stability over time, coverage, leakage against the decision date, fairness across segments, and value in dollars against what the client uses today.
Signals that pass every test are registered, versioned and allowed into a decision. The rest are dropped, with the reason kept.
| Signal | Lift | Stable | Coverage | No leakage | Fair | Value | Verdict |
|---|---|---|---|---|---|---|---|
| Buffer thinning against own normal | pass | pass | pass | pass | pass | pass | Kept |
| ATO payments slipping | pass | pass | pass | pass | pass | pass | Kept |
| Customers concentrating | pass | fail | pass | pass | pass | not run | Dropped |
| Next month's balance | pass | pass | pass | fail | not run | not run | Dropped |
| Weekend spending share | pass | pass | pass | pass | fail | not run | Dropped |
Illustrative example. A signal reaches a live decision only after it has passed every test.
Seven capabilities, on registered signals
Every use case draws on the same capabilities, so a new one starts from what already works.
Why did this change? A rise in arrears, split into who we lent to and how they performed.
Early arrears rose from 2.10% to 3.35%
Percentage points of the change. The two parts add up exactly.
Where the mix moved
| Segment | Share Q1 | Share Q2 |
|---|---|---|
| Loans $20k to $50k | 31% | 44% |
| Trades and construction | 18% | 26% |
| Under 2 years trading | 12% | 15% |
Illustrative example.
What is likely next? Delinquency forecasting for a renewal book: the forecast with its range, the drivers, and why this method suits this data.
Once live, actual delinquency is checked against the range every month.
Top drivers
Method chosen
A readable scorecard on changes against each customer's own baseline, because on held out renewals a larger model did not beat it.
Illustrative example.
What if? The book under each scenario, side by side, with ranges.
| Scenario | Arrears | 90% range | Losses |
|---|---|---|---|
| Base case | 2.9% | 2.5 to 3.4% | $4.1m |
| Cash rate up 1 point | 3.6% | 3.1 to 4.2% | $5.2m |
| Construction downturn | 4.4% | 3.7 to 5.3% | $6.8m |
Illustrative example.
What should we do? Choose the objective, set the limits, see the trade off.
Objective
Constraints
| Approval rate | at least 55% |
| Expected loss | at most 1.8% |
| Exposure per sector | at most 25% |
Illustrative example.
Is it worth it? The new approach against today's, on decisions already made, in dollars.
| Today's rule | New approach | |
|---|---|---|
| Loans that would change | n/a | 212 |
| Losses avoided | n/a | $640,000 |
| Margin forgone | n/a | $215,000 |
| Net | $425,000 |
Read with care
Only loans the lender actually wrote can be judged, because only they have outcomes. Declined applicants are left out, not guessed.
Illustrative example.
Is it still working? Predictions against outcomes, with an alert when they drift apart.
Illustrative example.
Ask in plain English. Every figure in the answer links to its evidence.
Why did arrears rise last quarter?
Most of the rise came from who we lent to: loans of $20k to $50k grew from 31% to 44% of new lending, and that group runs higher arrears. How borrowers performed added 0.35 points.
Each linked figure opens the calculation it came from.
Illustrative example.
The decision record
What every capability did is kept with the decision, so it can be explained, traced and contested later. How trust is built in
- The data snapshotThe rows used, as of the decision date, fingerprinted. Designed to be sealed so it can be proven later.
- The signalsEach figure's definition and version, traced to the raw data.
- The model and the policyThe model in force, with who approved it. The policy version is designed to be recorded beside it.
- The decisionInputs, outputs, reasons and uncertainty. Overrides are designed to carry their justification.
- The narrativeWhat the AI wrote, with every figure checked against the evidence.
- The outcomeWhat actually happened, joined back to the decision.
- ReplayDesigned to run the decision again later and give the same answer.
Illustrative example. Harbour Joinery is a synthetic business.
Proven before it is deployed
Every change follows one loop. Nothing is deployed until it beats what you do today, on your own history, inside your environment.
The same engine, other data
Comparing buildings with peers that operate the same way cut August false alerts at schools from 106 to 21. The same platform is applied in energy and retail, by conversation.
As AI agents start to draft memos and take routine actions, every number they use needs a source. That is what the engine is built around.
Book a conversation