From platform data to credit decision, in under 200ms.
A detailed look at how Lendforge ingests your signals, calibrates a model to your platform population, and delivers compliant credit decisions at the point of application.
The process
Five phases from data connection to live decisions
Connect via REST webhook stream or bulk CSV import. Lendforge accepts transaction records, user lifecycle events, product category data, and existing credit outcomes if available.
{
"event_type": "transaction",
"user_id": "usr_4Xk29mN",
"amount": 1240.00,
"category": "electronics"
}Lendforge normalizes raw events into model features: 90-day GMV, rolling velocity ratio, category concentration index, tenure percentile, repeat transaction rate, and 40 additional platform-native signals. Features are computed against configurable lookback windows (30, 60, 90 days) so the model captures both recent behavior and longer-term patterns. You do not manage feature engineering. Lendforge runs the pipeline and keeps the feature schema versioned alongside each model release.
A gradient-boosted ensemble is calibrated to your platform population. If you have historical outcome data (funded loans, repayment status, charge-offs), the model trains directly on your own defaults and survivorship patterns. If your program is new, Lendforge applies transfer learning from comparable population cohorts: similar platform type, transaction volume range, and product category mix. The calibrated model is specific to your population, not a generic consumer credit model applied to marketplace users.
Credit policy rules layer on top of model output. You define cutoffs, product tiers, and eligibility constraints. Lendforge enforces them deterministically before returning a decision.
POST /v1/decisions with application payload. Sub-200ms response. Synchronous approval, counter-offer, or decline with reason codes. Adverse action notices generated automatically for declined applications.
Model transparency
You should be able to explain every decision to a regulator.
Lendforge generates SHAP-based feature attribution for every decision. The contribution of each input signal, platform tenure, GMV trend, bureau tradelines, and the rest, is calculated individually and attached to the decision record. Your compliance team can see which signals drove a specific outcome, export the attribution to CSV for exam preparation, and build adverse action notices that cite the actual factors in plain language an applicant can understand and act on. When a state examiner asks why an application from June was declined, you have the exact model version, feature values, and policy rules that were active at that moment, not a reconstruction or approximation.
Integration timeline
Live in two weeks, not two months.
Week 1
Webhook or CSV data connector setup. Sandbox API keys and SDK install. Your engineers run end-to-end test decisions against a pre-built synthetic population of 10,000 applicant profiles spanning the credit spectrum, including edge cases for adverse action output.
Week 2
Model calibration on your historical platform data. Credit policy rules configuration: you define cutoff thresholds, product tiers, and exclusion logic. Compliance team reviews adverse action notice format and reason code mapping before any real applications are processed.
Go-live
Production API keys issued. Monitoring dashboard active. SLA clock starts. First real credit applications process through the live model. Decision audit log begins retention.