Acc‑01 · Healthcare
Interop
8 wk
Acc‑02 · Fintech
Risk Signal
6 wk
Acc‑03 · Salesforce
Agent
4 wk
Acc‑04 · Industry
Asset
10 wk
Risk Signal Accelerator · Fintech

From raw events
to a calibrated risk
score in six weeks.

A pre‑built risk scoring stack — feature store, model registry, scoring API, calibration dashboard, and the PSI monitor that catches drift before your auditors do. Tuned on your data in week one, in production with a real backtest by week six.

Book a pilot scope
Latest: v3.1.0, Feb 2026 License: Per deployment

What you actually
get on day one.

9 modules · git‑cloneable
M.01

→ Feature store

Online + offline feature parity. 200+ pre-built features across velocity, recency, monetary, device, and graph dimensions. PIT-correct.

Feast · Redis · Snowflake
M.02

→ Model registry

Versioned models with reproducible training runs. Model cards auto-generated from each run. SHA-pinned dependencies.

MLflow · Git LFS
M.03

→ Scoring API

REST + gRPC endpoints with idempotency keys, retry-safe semantics, and per-model A/B routing. p99 < 100ms at 1.8M/day.

FastAPI · gRPC · Envoy
M.04

→ Calibration dashboard

Bucket-wise calibration plot, Brier score, expected calibration error. Your risk team owns it. We hand it over week three.

Streamlit · scikit-calibration
M.05

→ PSI / drift monitor

Weekly Population Stability Index per feature + score. KS test against baseline. PagerDuty alert at PSI > 0.25.

Evidently · PagerDuty
M.06

→ Backtest harness

Replay any historical window against any model version. Vintage analysis, lift curves, ROC at custom thresholds.

Polars · DuckDB
M.07

→ Explainability

SHAP per-score explanations served with the score. Aggregate feature importance dashboard for the risk committee.

SHAP · TreeExplainer
M.08

→ Compliance reporter

Model documentation, decision logs, adverse action codes, audit trail. RBI-aligned reporting templates included.

RBI · IFRS 9 · PDF gen
M.09

→ Runbook + SLA

17 named runbooks for model incidents. 99.95% uptime SLA on Hosted tier. Drift escalation paths pre-defined.

Runbook v3.1 · SLA-backed

The scoring pipeline,
flattened out.

5 stages · < 70ms p99

Five stages, two stores, one source of truth.

risk-signal-accelerator v3.1.0
EVENT TX STREAM STAGE 1 Enrich Event ~8ms STAGE 2 Build Features ~22ms STAGE 3 Score (XGBoost) ~14ms · core STAGE 4 Calibrate ~6ms STAGE 5 Decide + Explain ~12ms · SHAP feature store decision log · every score
Stage 1
Enrich Event
Resolve customer, account, device, merchant. Idempotency keyed.
Resolver · MDM
Stage 2
Build Features
200+ canonical features. Online via Redis, PIT-correct via Snowflake.
Feast · Redis
Stage 3
Score
XGBoost / CatBoost ensemble. A/B router for champion-challenger.
XGBoost · ONNX
Stage 4
Calibrate
Isotonic + Platt calibration. Probability → action threshold.
scikit-calibration
Stage 5
Decide + Explain
Decision + SHAP explanation + adverse-action code, in one response.
SHAP · TreeExplainer

The six‑week plan.
Pinned, not promised.

6 weeks · 3 engineers
Week ↓Ship
Week 1
Data audit + baseline.
Two engineers virtual with your risk and data teams. Catalogue available signals, PII boundaries, and ground-truth labelling. Baseline model trained on day five.
Baseline AUC
Week 2
Feature store live.
Online + offline feature parity verified. 200 canonical features hydrated from your data warehouse. PIT-correct backtest harness running.
Feature store
Week 3
First production-class model.
XGBoost ensemble trained on your data. Calibration dashboard handed to your risk team. Model card and reproducible run published.
Champion model
Week 4
Scoring API in shadow.
Live scoring in shadow mode against production traffic. Latency benchmarked. Explainability + adverse-action codes wired in.
Shadow live
Week 5
Backtest + sign-off.
12-month backtest with lift curves, vintage analysis, capture rate at custom thresholds. Risk committee sign-off meeting.
Backtest report
Week 6
Production cutover.
Production cutover at 5% → 100% over three days. PSI monitor on. RBI-aligned compliance pack delivered. 30-day safety net begins.
In production

What our last
three customers said.

Verified · 2025—2026
"
Our previous engine fired on too many good transactions and missed the coordinated mule networks entirely. The Risk Signal Accelerator gave us 94% recall at sub-percent FPR in six weeks — work that took our previous vendor two years.
Head of Fraud, Top-3 Indian Private Bank
1.8M tx/day · live since Jan 2026
Time saved
22months
Recall lift
+18pts
FPR cut
−87%

Risk score
in five weeks?

↳ Direct line · Risk Signal Accelerator

Tell us your use case, your scale, and your regulatory posture. We'll send a pilot scope, architecture sketch, and fixed-fee quote within three business days.

Get in Touch

Questions we
get every week.

Click to expand ↓
Q.01
Will you train on our PII or push it to the cloud?+
No. PII never leaves your perimeter on the Pilot tier. On Hosted, we run inside a customer-owned VPC. All features use hashed identifiers. Raw PII columns are explicitly excluded by the feature catalogue.
Q.02
What use cases does the Accelerator cover?+
Card fraud, account-takeover, mule networks, credit risk (PD), AML transaction monitoring, and merchant onboarding risk. Each ships with a baseline model tuned for the use case. Custom use cases plug into the same pipeline.
Q.03
How do you handle the RBI model-risk circular?+
The compliance reporter generates the documentation the RBI circular asks for: model purpose, methodology, validation, monitoring plan, and material change log. We've passed two RBI inspections on it.
Q.04
Champion-challenger — how does that work?+
The scoring API routes a configurable share of traffic to a challenger model. Outcomes are logged and compared on a recurring statistical-significance test. Promotion is gated by your risk team, never automatic.
Q.05
Can it score in real-time at our scale?+
Yes. The reference architecture sustains 1.8M tx/day at p99 < 100ms. We've stress-tested to 4× that ceiling. Scoring is horizontally scalable on K8s.
Q.06
What happens after we stop paying?+
On Pilot or Enterprise, the source is yours forever. You run it. On Hosted, we ramp down with a 90-day handover. We've never lock-in'd a customer in nine deployments.