PROBLEM × SOLUTION
Five pain points. Five capabilities.
Side by side.
Every credit operation carries the same paradox: too conservative loses good clients, too loose grows defaults. BlueRisk hits each side with a direct capability, same order, same weight.
WHAT HURTS TODAY
WHAT BLUERISK DOES
Outdated and generic model
The market score you license was trained on your 2019 book. The client has changed. The model hasn't.
Slow, over-conservative decisions
A conservative policy lowers defaults in the short term but throttles approvals. Good clients go to the competitor.
High unit cost on review
Manual review at fintech volume turns into a credit-desk queue. Each review costs hours, and the good cases leave with the bad.
Growing without growing risk
The base grows, defaults grow. The operation hits a ceiling: either slow the growth or accept the loss.
Decisions with no explanation for the regulator
BACEN and LGPD ask for the reason. When the model is a black box, the only defense is "because we said so," and that no longer flies.
Score tuned to your book
The model trains on your data, your client profile, your events. It doesn't compete with the market score. It coexists with it.
Automated decisions with human-in-the-loop
Clear cases (green and red) ship in seconds. The grey, and only the grey, goes to human review with the summary ready.
Cost per decision that drops with automation
Routine review leaves the back office. Analysts go back to what humans do best: the edge case, the exception, the negotiation.
Sources beyond the bureau
Behavior on your base, Pix transactions, public data, anti-fraud signals. The score reflects the real client, not just the bureau history.
Auditable by design (BACEN, LGPD)
Every decision records the input, the per-feature weight, the cutoff reason, and the model version. Native White Box AI, not an overlay.
USE CASES
Where BlueRisk lands first.
Four recurring fronts where the gain shows up in weeks, not quarters.
01
Credit score
Proprietary model, modern ML, explainability on every request.
02
Automated underwriting
Policy run as code. Instant approval for what fits, human review for what doesn't.
03
Classification and segmentation
Book segmented by risk, with offers and treatments calibrated per band.
04
Fraud detection
Pix, card, identity. Anomaly in milliseconds, with handoff to the anti-fraud desk.
TYPICAL IMPACT
The numbers the pack
tends to deliver.
Ranges observed across BlueRisk projects in production. They vary by product and book maturity.
ARCHITECTURE
Four technical capabilities.
One platform.
Modern models calibrated on your book, with native explainability for the regulator. Grey cases go to a human with the context ready, not with a PDF to read.
Score
PD (probability of default) and LGD models calibrated on your book, with SHAP explainability per feature.
- modern models
- SHAP explainability
- vintage versioning
live preview
Score
Decision
Policy run as code, with a rule engine and ML running in parallel. Audited human override.
- rules + ML
- logged overrides
- circuit breakers
live preview
Decision
Classification
Segmentation by behavior and risk. Offers calibrated, treatment matched to each band.
- supervised clustering
- time-series behavior
- evolving segmentation
live preview
Classification
Detection
Transaction or identity anomaly in milliseconds. The signal goes to the anti-fraud desk with context.
- sub-100 ms inference
- online features
- handoff with context
live preview
Detection
HOW IT WORKS IN RUNTIME
From request to decision, in minutes.
Request arrives
application, transaction, or onboarding. Internal data, bureau, and external signals.
Model decides
score, policy, and detection run in parallel. The decision ships with its reason.
Human reviews the grey
only what falls outside the guardrail goes to the desk, with the summary ready.
Decision logs and learns
structured output to the core. The real outcome feeds back, and the model recalibrates.
ENGAGEMENT
Three phases.
Gradual entry, clear cycle.
Not an investment table, it's the project journey. It starts with a single policy and expands on evidence.
Pilot
5-7 weeks
1 priority policy, baseline score, bureau integration, baseline metrics.
Growth
10-14 weeks
Multiple products, fraud detection in production, fully automated decisioning on at least one flow.
Operation
Recurring
Vintage recalibration, continuous monitoring, new models, regulatory evolution.
FOR WHOM
TECH STACK
Built on recognized foundations.
Amazon SageMaker
Training, inference, and governance of predictive models at scale, with vintage-level versioning.
AWS Bedrock
Frontier models for case summarization, reason generation, and analyst support.
Databricks
Lakehouse for risk data (internal, bureau, and external signals), with a semantic layer and lineage.
COVERAGE
Cases covered by the pack.
Pre-trained models covering the most common Brazilian credit flows. Specific variations are calibrated on your book during the Pilot phase.
Underwriting
- PD score
- LGD
- EAD
- Behavioral score
- Propensity to contract
- Dynamic limit
Monitoring
- Early warning
- Vintage score
- Action recommendation
- Credit churn
- Retention
- Reset
Fraud
- Real-time Pix
- Card present/not-present
- Identity
- AML
- Money mule
- Ring patterns
Collections
- Propensity to pay
- Right tone
- Preferred channel
- Recommended deal
- Expected recovery
CLIENTS USING IT TODAY
Real results, in production.
DIRECIONAL
An ML risk model cut default rates nearly in half at a Brazilian real-estate developer. Decisions with objective criteria, standardized and auditable.
FREQUENTLY ASKED
FAQ
The regulator accepts AI in the decision as long as every case can be explained and audited. BlueRisk uses models with native explainability (SHAP, cutoff reason, model version, input data) and keeps the record for the regulatory retention period. Policy as code makes supervisory review easier.
Yes. The pilot always starts with a single policy, usually the highest-volume and highest-pain one. Score, decisioning, and monitoring go to production in that product before expanding to the others.
Internal and bureau data stay in your environment (your AWS account, or BlueMetrics' AWS account under your control). The model never exports client data. Logs, training, and inference run inside the perimeter you define.
Every decision reaches the analyst with the reason (per-feature weight), the borrower's score history, a comparison to the equivalent band in the book, and suggested actions. Human overrides are logged and feed back so the model keeps learning.
For PD/LGD in standard credit, 24 months minimum, ideally 36. For transactional fraud, 6-12 months is enough: fraud changes fast, so the model needs short recalibration cycles. The pilot is designed around the real history you have available.
Yes. Real-time Pix, card present and not-present, identity fraud, AML: all share the same scoring and detection architecture. For unlabeled projects (new fraud), we use clustering and anomaly detection before the supervised models come in.
BEYOND THE PACK
Model needs proprietary vertical
architecture? That calls for a custom project.
When the problem doesn't fit the Solution Pack frame, Custom Projects take over: open scope, a senior multidisciplinary team, the same engineering standard.