Lifetime Credit Risk Platform

Underwriting across
the life of the loan,
with reasoning attached

New underwriting, monitoring of existing loans, additional-loan assessment and model validation, all on one platform. Instead of a single PD at drawdown, it produces a risk curve over the life of the loan, and every decision leaves a written rationale and a record of who confirmed it. Your existing credit systems stay as they are.

Request a demo →Talk to us about a rollout
Term-structured
PD curve
Cumulative and conditional, N months after drawdown
Three-band
decision routing
Fast-track, review, deep review — never an automatic decline
Every decision
carries a written rationale
From SHAP, on the same axis as the score
1.8e-7
Frozen-build reproduction deviation
Seeds fixed · frozen-lineage audit
View all products →
THE PROBLEM

What underwriting today
cannot answer

Underwriting that ends
at a single score
A credit score and a point-in-time PD never say when. They cannot see that the risk piles up right as the grace period ends.
A decline you cannot explain
Korea's AI Framework Act classifies credit scoring and loan underwriting as high-impact AI. The reason for a decision has to be given in sentences — and on the same axis that produced the score.
Customers with thin files
A first-time small business or sole proprietor has empty fields. Treat the gaps as deductions and you shut out good customers; ignore them and you miss real risk.
ARCHITECTURE

Three layers

Layer 3 — screens, approval, integration
Confirmation gate + connectors to your existing credit systems
New underwriting · loan monitoring · model validation · insightDecision criteria · change history · recorded override reasons
Chat assistant · printable reportREST API · integration with your credit systems
Layer 2 — the decision layer
Term-structured risk curve · grade · three-band routing
lifetime PD term structureFine-grained grades 1–10Three-band routing
Early warning signalsConfidence adjustment for thin filesPD → terms policy simulation (limit and rate what-ifs)
Layer 1 — the model layer (Syndasis Core ③ quant & predictive / ④ explainable decision)
Survival analysis + ensemble · calibration · explanation · disclosure
Discrete-time survival analysis + a cross-sectional XGBoost ensembleCalibration (binomial test by grade · Hosmer–Lemeshow)Natural-language rationale from SHAP
Fairness (DP · EO)Drift (PSI · CSI)Frozen-lineage audit
6 WORKFLOWS

Six working capabilities

CapabilityWhat it doesThe question it answers
New loan underwritingScores an application in real time, producing a 12-month PD, a curve over the life of the loan, a fine-grained grade and a written rationale, then routes it into one of three bandsShould we approve this one?
Monitoring existing loansRecalculates the remaining conditional risk on live loans and sorts the review queue by grace-period expiry and early warning signalsWhat should we look at first?
Additional-loan assessmentCarries over repayment and delinquency history from the original account to recalculate risk on the additional exposure, with what-if comparison of changed termsCan we lend more, now?
Model validationShows discrimination, calibration, stability and fairness metrics with a three-line reading of each, plus performance over timeCan we trust this model?
Insight and decision criteriaPortfolio findings, thresholds and grade boundaries with their change history, and what external data is in use — on one screenWhat are the criteria, and when did they change?
Chat assistantAn in-house LLM (deployable on a closed network) that answers from the values shown on screen. If a value isn't on the screen, it says soWhat does this number mean?
HOW IT RUNS

How it works

A hybrid model
Time structure → discrete-time survival analysis
Cross-sectional discrimination → XGBoost
Combined → a calibrated ensemble
The LLM is used only to check rationale sentences and to power the chat
Reasoning on every decision
The top three contributing factors from SHAP
Three to five sentences of natural-language rationale
A confidence segment per decision
LLM-written sentences are shown only after they pass post-hoc verification
Three levels of underwriting autonomy
Level 1 — recommendation only; a person confirms every band
Level 2 — streamlined review in the fast-track band
Level 3 — terms assigned from policy simulation
A decline is confirmed by a person at every level
Data principles
The same calibration table in training and serving
A missing value is not a deduction
One source of truth for the default definition, in a config file (swappable per institution)
Out-of-time splits only
Validation and audit
Every random seed fixed · frozen-build lineage audit
A sealed holdout kept separately
Monthly PSI monitoring · retraining triggers
A defined procedure when a fairness metric moves
How it is rolled out
A standard credit schema adapter (accounts, repayments, delinquency, company data)
Per-institution configuration (default definition, thresholds, grade boundaries, protected attributes)
SaaS · closed network · hybrid
COMPLIANCE

High-impact AI requirements,
met on screen, not in a binder

Korea's AI Framework Act classifies credit scoring and loan underwriting as high-impact AI. Rather than producing separate artifacts, the requirements are met inside the screen an underwriter uses every day. The right to explanation under the Credit Information Act is answered by the same rationale sentences.

Risk management
Identify · assess · mitigate, as a matrix
Model risk, data risk, fairness risk and operational risk are identified item by item and managed as a matrix of assessments and mitigations. Drift metrics and retraining triggers are wired directly into those mitigations.
Explanation
The outcome · the criteria · an overview of the training data
The decision, the probability and grade, the thresholds and grade boundaries applied, and the scope and period of the training data are all visible on screen. The rationale sentences come from the same SHAP computation that produced the score.
Human oversight
Marked as a recommendation · overrides recorded
Every output is labeled a recommendation, and a person confirms it. A decision that differs from the system's is recorded with its reason and remains auditable alongside the history of criteria changes.
Prior notice
Notice language for the use of AI
Ships with notice language telling the applicant that AI is used, and a decision-notice template that carries the rationale sentences. Both can be replaced with your own wording.
AI judges, a person decides, and your existing credit systems execute.
There is no automatic decline. The platform lightens the review load by splitting it into three bands; every final call, declines included, is made by a person.
SCENARIOS

Rollout scenarios

Policy finance · guarantee institutions
A first-time small-business application arrives
→Fields with no history are treated as a confidence adjustment, not a deduction, so thin-file applicants aren't systematically underrated
→A curve over time and a grade are produced, and a cost-weighted threshold routes the case into the fast-track band.
→An underwriter confirms it through the streamlined review, and a notice containing the rationale goes out to the applicant
Savings banks · mutual finance monitoring
You have to decide which live loans to review first
→Loans three months from the end of their grace period are selected automatically, and the conditional risk over the remaining term is recalculated
→Signals such as a change in repayment pattern Early warning signals are shown alongside, so you can tell what is driving the increase
→It moves into a prioritized review queue, and the outcome of the review is kept as history
Fintech lending platforms
An existing customer applies for an additional loan
→Repayment and delinquency history from the original account is carried over and the risk after the additional exposure is recalculated
→How risk moves when the limit, rate or maturity changes is compared as a what-if.
→The policy layer proposes candidate terms and a person confirms them. The reason for the decision is recorded with it

See whether your credit data
can support a first model

We have an adapter for a standard credit schema — accounts, repayments, delinquency, company data — and a per-institution configuration layer.
Tell us what data you hold and we will work through feasibility and scope with you.

Request a demo →