Flexible integration
Two read-only paths—standard API and file upload—with no rebuild and no write-back to in-house systems. Most customers complete validation-stage analysis on file upload alone.
Four Capabilities · Four Agents
Four capability Agents forming the hub for risk analysis and verification
Strategy Sandbox, Model Diagnosis, Rejected Applicant Recall, and Feature Value Analysis—four independent Agents sharing one data foundation. Every conclusion carries its rationale, and every detail can be traced layer by layer, so risk judgment is always backed by evidence.
Capability Overview
Four capabilities, four kinds of problems, one foundation
The four capabilities may look independent, but they share the same data foundation and analysis engine, forming an interconnected capability chain: validation conclusions from Strategy Sandbox point to Model Diagnosis, diagnostic findings feed back into Feature Value Analysis, and gray-release results from Rejected Applicant Recall flow back as new validation samples.
| Capability | Use case | Input | Output |
|---|---|---|---|
| Strategy Sandbox | Ready to adjust a strategy, but unable to predict the impact | Historical decision data, rules / thresholds to adjust | Post-adjustment false-decline rate, miss rate, affected segments, and rule risk alerts |
| Model Diagnosis | Model performance is slipping, but the cause is hard to locate | Model input inventory, decision outcomes, and performance labels | Decay attribution conclusions, failed-feature inventory, alternatives, and a reproducible evidence package |
| Rejected Applicant Recall | False declines suspected, but hesitant to loosen the approval threshold | Historical decline records and high-quality segment performance data | Similarity matching results, recall effectiveness analysis, recallable list, and recall recommendation report |
| Feature Value Analysis | At data procurement renewal, but the actual value is hard to quantify | Invocation logs and model input inventory | Data source value metrics table (KS / IV / PSI / cost / lift) and low-value source inventory |
The problem it solves
Before go-live, run simulation backtests on historical data to quantify the false-decline and miss changes caused by threshold or rule adjustments, and decide whether to ship on that basis; rule conflicts are scanned automatically, proactively identifying and flagging rules with logical conflicts, mutual overshadowing, or long-term zero hits.
How it works
01
Scope the backtest Select the strategy pool and time window; extract real decision data and performance labels.
02
Define the adjustment Adjust thresholds and add or remove rules in the UI, with multiple plans compared in parallel.
03
Run the backtest Replay the new strategy over historical data to project its outcomes, compressing the evaluation cycle from days to hours.
04
Read results with evidence Compare metrics such as misses, false declines, interception lift, and rule hits—each metric can be expanded layer by layer to show its basis.
05
Detect conflicts Scan relationships between rules and flag those with logical contradictions, overshadowing, or long-term zero hits.
What it tells you
If this rule’s threshold moves from 0.7 to 0.65, the Strategy Sandbox quantifies the resulting false-decline and miss rate changes for the strategy pool and identifies the structural profile of the affected segments; it also flags two logically conflicting rules in the pool—one declining, the other approving—one of which over-declines high-quality customers, already marked red in the corner; three more rules have had zero hits in six months and are recommended for retirement.
Strategy Sandbox Agent Runs backtests, comparisons, and conflict detection around the impact of strategy adjustments, producing validation conclusions backed by evidence.
The problem it solves
When model performance slips, it automatically builds a performance baseline and attributes the decay at feature level—pinpointing whether it stems from feature drift, segment shifts, or data-source degradation—then delivers a reproducible fix with expected gains, eliminating round after round of trial and error.
How it works
01
Build the baseline Compute current true performance, compare against the historical baseline, and quantify the performance decay.
02
Attribute at feature level Evaluate each feature’s distribution drift, discriminative power, missing rate, and correlation to locate the features dragging the model down.
03
Validate by cohort Break results down by segment to identify whether the problem is global or concentrated in specific cohorts.
04
Run ablation experiments Quantify the performance change after removing a feature—confirming attribution by experiment, not conjecture.
05
Recommend the fix Provide alternative feature plans with expected effects, exportable and reproducible on-premises.
What it tells you
This round of model decay was driven mainly by drift in the “device score” feature, caused by iOS version changes over the past two months. The platform replaces it with “behavioral biometrics”, recovering performance to a quantified level—a conclusion verified by a real experiment and open to review. Every conclusion ships with three layers of evidence: feature-ablation experiments, cohort validation, and raw records. The evidence package exports and reproduces 100% on the customer’s premises.
Model Diagnosis Agent Builds a reviewable attribution chain for model decay from performance baselines, feature-ablation experiments, and cohort validation.
The problem it solves
Historical similarity matching brings high-quality customers who were falsely declined back into credit—real historical performance replaces prediction, producing verifiable, back-stoppable recall lists and risk references, so the business team dares to loosen approval and can still pull back steadily when risk rises.
How it works
01
Scope candidate declines From risk-control screening results, select “borderline declines”—applications rejected by rules but without strong risk signals.
02
Identify segments Segment the borderline declines, clarifying the dimensions on which they resemble high-quality customers.
03
Match against history Run multi-dimensional similarity matching against historically approved, well-performing customers. Every number comes from historical observation—observed values replace inference and estimation.
04
Analyze recall outcomes For each batch of candidates, produce the similarity distribution, recallable volume, and risk reference.
05
Issue the recall report Output the recallable list and risk reference report with matching rationale and applicable boundaries, for business staff to judge and confirm.
What it tells you
This declined application shares over 0.85 similarity with 1,247 high-quality customers, whose actual historical delinquency rate is 4.2%—an observed value, not a prediction. The system produces the recallable list and risk reference report, each with matching rationale and applicable boundaries. Every number presented comes from real observation in historical data—observation replaces inference, estimation, and black-box judgments, and the conclusions stand up to checking.
Rejected Applicant Recall Agent Builds verifiable, traceable recall lists and risk references on the basis of real historical performance and similarity matching.
The problem it solves
Connect each data source’s invocation logs with the model input inventory, then compute KS / IV / PSI / cost / lift per source—flagging high-cost, low-lift sources, so every procurement investment has a measurable return.
How it works
01
Connect logs with the input inventory Establish which features each source contributes and whether they are actually used in the model.
02
Aggregate usage by source Count features provided, features actually used, and the utilization rate per source.
03
Compute value metrics per source KS / IV (discriminative power), PSI (stability), and lift (incremental model contribution)—all computed from real invocation logs and the input inventory.
04
Generate the value tracking table KS / IV / PSI / cost / lift for every source at a glance, refreshed monthly.
05
Recommend trade-offs Flag high-cost, low-lift sources and quantify the model loss if they are cut.
What it tells you
Which data sources truly deliver lift, which have contributed little for years while still being paid for, and which two overlap so heavily that keeping one suffices. How much model accuracy is lost after retiring a source—a figure computed from real invocation logs and the input inventory, backed by data. In procurement negotiations, the business team finally gains a quantified basis for decisions.
Feature Value Analysis Agent Brings invocations, model usage, performance, and cost into one value chain, continuously accumulating quantifiable evidence for data procurement and governance.
Use Cases
Four capabilities covering the full business chain
The four capabilities embed across the full business chain, delivering continuously at every point of the front line—so the risk chain grows stronger with repeated refinement.
Strategy Sandbox Validate strategy impact before go-live
Credit application anti-fraud
When new fraud patterns emerge, see the full impact of an adjustment in the sandbox before deciding whether to ship it.
Credit approval rate optimization
Rehearse borderline approvals before go-live and see the trade-off between approval rate and risk.
Full model lifecycle
Validate with sandbox simulation before launch—no blind tuning.
Model Diagnosis Decay localization & repair evidence
Credit model governance
When a scorecard decays, locate the root cause and deliver reproducible repair evidence—no repeated trial and error.
Full model lifecycle
Monitor continuously after launch; decay is caught the moment it appears.
Rejected Applicant Recall Real matching for borderline segments
Credit application anti-fraud
Falsely declined customers are safely brought back through similarity matching and recall validation.
Credit approval rate optimization
Bring back borderline segments worth serving—from “rather over-decline” to “dare to approve, able to pull back.”
Feature Value Analysis Data contribution & procurement governance
Data source value assessment
Regularly compute each source’s true contribution and cost, grounding procurement negotiations in data—so every yuan spent on data procurement has a clear return.
See the capabilities in action Every business context differs. Upload desensitized data and run one product demo to see how the capabilities land in practice.
Start a Product Demo → Foundation & Deployment
Appliance deployment—data never leaves your environment
A single on-premises appliance is the unified carrier: it connects in-house data with the existing risk-control stack and, through the unified data foundation, analytical capabilities, and Agent runtime, powers four kinds of intelligent risk applications. The appliance runs alongside your existing risk-control system, focused on analysis and verification—making what you already have stronger.
Continuous evolution, independently controllable
Three classes of upgrades—Agent capabilities, foundation models, and platform features—iterate continuously, making the platform stronger with use. Upgrades change only “how analysis is done”; your strategies, models, variables, and historical data remain independently held and exportable at any time.
Security & compliance
Independently isolated data, clear permissions, and fully audited processes—aligned with the security standards of your existing risk-control system.
One appliance carrying data and capabilities, making the existing risk-control system stronger.
Get Started
One product demo to see the real chain of risk analysis
With desensitized scenario data, experience the full “analyze—verify—recommend” chain: locate the problem, present the evidence, and clarify the next tuning direction.
The demo uses desensitized data and never touches your production systems.