Centurion · Decisioning · FlowX
Where credit and fraud signals converge.
The production decisioning fabric. Where the model meets the customer, the bureau pull becomes a tradeline, and the strategy becomes a decision. Not a generic rules engine — purpose-built for the decisions money is made and lost on.
Tradeline parsing, real-time scoring, A/B testing, adverse-action reason codes, audit-grade logging — every primitive of credit and fraud decisioning, in one fabric.
Why purpose-built
A generic rules engine fires on attributes. It cannot make a decision.
Built by practitioners who have run real-time decisioning at the scale where milliseconds and dollars are measured together — and codified the playbook they ran there.
The people behind FlowX have operated credit decisioning at tier-1 card issuers, run fraud strategy at money-center banks, and built the real-time decisioning stacks at fintech lenders through every stage from launch to securitization. We've lived inside the audit conversations, the override workflows, and the late-night incident reviews that come with decisions of consequence.
Generic rules engines and BPM workflow tools fire on attributes you happen to have. They don't parse tradelines. They don't treat models as versioned, A/B-testable artifacts. They don't produce adverse-action reason codes natively. Their audit trail is something you build on top. FlowX treats all of these as first-class primitives — because in credit and fraud, they aren't optional.
Anatomy of a decision
What happens in the milliseconds before a decision lands.
Every FlowX decision is the composition of nine stages. Each one is explicit, versioned, observable, and replayable — not buried inside a rule tree that grew over five years.
- 01
Bureau intake
TransUnion, Experian, Equifax payloads ingested, snapshotted, pinned to the decision.
- 02
Tradeline parsing
TU TLS, Experian CDF, Equifax ARF — parsed into the consumer/tradeline hierarchy.
- 03
Attribute computation
Thousands of bureau attributes computed in real time. Same library Model Foundry trained on.
- 04
Segmentation
Route to the right model and policy set — thin-file, thick-file, derog, NTC, channel-specific.
- 05
Model scoring
PD, fraud, behavioral, collections — versioned, logged, pinned to the strategy version.
- 06
Real-time graph signals
Native ML Graph call — entity resolution, mule proximity, ring detection — at decision latency.
- 07
Lists & policy knockouts
Negative lists, allowlists, sanctions, MLA/BK/KYC cutoffs. Hard rules override scores when policy demands.
- 08
Strategy execution
Cutoffs, tier assignment, line size, treatment routing — versioned and A/B-testable.
- 09
Reason-code attribution
FCRA §615 adverse-action reasons attributed, ranked, mapped to consumer-readable language.
- 10
Decision log
Every input, score, rule firing, and reason — written to an immutable, replayable log.
Bureau-native
Tradeline-level. Hierarchical. Real-time.
FlowX speaks the credit bureaus' native languages. Pulls, parses, attributes — out of the box. The same hierarchy and trended history that Model Foundry trains on, available at the moment the decision is being made. No flattening. No information loss between training and production.
Bureau intake
Ingests pre-pulled bureau payloads from your existing pull layer — versioned, snapshotted, and pinned to the decision record for replay.
Tradeline parsing
TU TLS, Experian CDF, Equifax ARF — natively parsed into the consumer-header / tradeline / inquiry / public-record / collection hierarchy.
Attribute library
Thousands of bureau attributes — CreditVision, Premier, Trended 3D, ATB — computable in real time on the parsed structure.
Custom attributes
Define and version institution-specific attributes alongside bureau-standard ones. Same lineage. Same documentation. Same audit trail.
Dev / prod parity
The attribute library used in production is the same one Model Foundry trains on. The 680 score in development is the 680 score in production.
Point-in-time replay
Re-run any historical decision with the exact bureau snapshot, attribute library version, and strategy that ran originally. For audit, for debugging, for back-testing.
Native signals & controls
Real-time graph signals and list controls — built in, not bolted on.
The two categories of signal that fraud teams need at decision time — network-aware graph scoring and operational list management — are first-class primitives in FlowX. Not external service calls. Not loosely-coupled microservices. Native, latency-budgeted, fully logged.
ML Graph integration
Real-time graph scoring, in the decision path.
FlowX calls ML Graph natively at decision time. The graph is updated as decisions are made — every new identifier, link, and event becomes part of the population the next decision is scored against. Detection compounds.
- Entity resolution across email, phone, device, address, payment instrument
- Mule-network proximity scoring (k-hop, weighted, time-decayed)
- Ring detection at population level — coordinated cohorts surfaced
- Shared-identifier overlap density between applicants and known-bad nodes
- Cross-portfolio scoring when multiple lines of business share the graph
List management
Negative lists, allowlists, sanctions — managed, versioned, auditable.
The operational controls that fraud teams actually use day-to-day — block lists, trusted-customer lists, sanctions screening — are first-class FlowX objects. Versioned, role-gated, audit-logged. Not a spreadsheet, not a side-channel service.
- Negative lists across device, IP, email, phone, BIN, IBAN, address
- Allowlists for trusted customers, internal accounts, integrated merchants
- Sanctions / OFAC / PEP screening with fuzzy-match scoring
- Industry consortium lists (shared-fraud, mule, identity-theft)
- Time-bounded, role-gated edits with full who-what-when audit trail
- Automatic expiry, review cadence, and decay logic per list
Strategy A/B testing
Champion / challenger as a first-class primitive — not a spreadsheet exercise.
Every strategy change in FlowX runs through a designed experiment: holdout, treatment, statistical power, ramp logic, and credible measurement of the lift on the outcomes that matter — approval rate, bad rate, override rate, profitability per booked account.
Designed holdouts
Cells defined ahead of the test, with power analysis on the metric of interest. Not p-hacked post-hoc.
Champion / challenger
Multiple challenger strategies in flight simultaneously. Promotion follows the data, not the meeting.
Shadow scoring
Score a candidate model against live traffic without changing the live decision. Catch regressions before they ship.
Decision-quality lift
Statistical AND economic comparison. Approval-rate delta, bad-rate delta, expected-loss delta, profitability delta — by segment and by vintage.
Ramp control
Begin at 1% traffic, ramp on observed performance, kill-switch at predefined guardrails. Production safety as a contract, not a wish.
Treatment effect attribution
Lift attribution that isolates the strategy change from concurrent macro and seasonal effects. The number you report is the number that's real.
Decisions powered
Every decision in the credit and fraud lifecycle.
One fabric. One audit log. One A/B framework. Whether the decision is to approve a loan, raise a line, decline a payment, route a collection call, or step up a transaction for verification.
Credit
- Application / origination decisioning
- Risk-based pricing tier assignment
- Line-assignment and credit-limit allocation
- Loan-amount and term selection
- Credit-line increase / decrease (CLI / CLD)
- Behavioral re-scoring & authorization strategy
- Collections treatment routing & contact strategy
- Recovery & post-charge-off routing
Fraud
- Account-opening decisioning (synthetic, identity, first-party)
- Login & session-trust scoring
- Payment / transaction authorization
- Insurance-claim decisioning
- Promotional-redemption decisioning
- Step-up & verification routing
- Mule & money-movement detection
- Manual-review queue prioritization
Every decision maps to one of four outcomes — APPROVE · PROTECT · STEP UP · BLOCK — with calibrated thresholds, reason-code attribution, and full audit trail.
Replay
Time-travel debugging. For audit, for backtest, for the incident at 3am.
Every decision FlowX makes is fully reconstructible — not just logged, but re-runnable with the exact bureau snapshot, attribute library version, model versions, graph state, lists, and strategy that ran the first time. Replay is a first-class primitive, not a forensic exercise.
Single-decision replay
Pull any historical decision by ID. Reproduce the inputs, attributes, scores, and rule firings exactly. The 3am incident review and the regulator's question both end with the same artifact.
Counterfactual replay
Replay the same decision with a different strategy version, model version, or cutoff. Quantify the impact of a proposed change on real historical traffic before shipping it.
Population backtest
Replay a candidate strategy across a defined population segment. Approval-rate delta, bad-rate delta, override delta — measured against ground truth, on actual decisions.
Point-in-time data fidelity
Bureau snapshots, attribute library versions, lists, and graph state are all versioned. The replay reflects the world as it was — not as it is today.
Dispute defense
When a consumer disputes a decision, replay produces the complete chain of inputs and reasons in minutes. The adverse-action notice and the replay artifact reconcile by construction.
Validation evidence
MRM and validators can replay any model in production against any historical window — without a special data pull, without a re-engineered fixture, without a multi-week project.
Audit by construction
Every decision, fully reconstructible. Years later. On demand.
The audit trail isn't a feature you turn on. It's how FlowX works. Every decision writes a complete record — inputs, attributes, scores, rules, reasons, strategy version, model versions — into an immutable log. When a regulator, an auditor, or a consumer asks why, the answer is two clicks away.
Immutable decision log
Every decision is logged with a full feature snapshot, strategy version, model versions, rule firings, and reason codes. Tamper-evident, queryable, exportable.
Point-in-time replay
Re-run any past decision with the exact data, attribute library, and strategy that produced it. For dispute defense, validation, or back-testing a counterfactual.
FCRA §615 adverse action
Reason codes attributed from score factors, ranked, and mapped to consumer-readable language. Auditable mapping from score to notice.
ECOA / Reg B disparate-impact
Per-decision protected-class proxy attribution (BISG / BIFSG), with aggregate disparate-impact reporting available on demand.
SR 11-7 alignment
Model inventory, version pinning, monitoring records, and ongoing-performance evidence — the artifacts the MRM function is built to consume.
Regulatory disclosure outputs
CFPB Section 1071, OCC examination requests, internal audit queries — the data is structured to answer them without a special project.
Production discipline
The things that fail at 3am, accounted for at design time.
Latency budgets
Per-strategy SLAs with circuit breakers and fallback paths. Slow scorers degrade gracefully, not catastrophically.
Instant rollback
Every strategy is a versioned artifact. One-click revert to the last known-good. No restart. No outage.
Override workflow
Manual reviewer interface with override tracking, override-driven retraining loop, and override-rate monitoring as a model-health signal.
Monitoring dashboards
Approval rate, bad rate by vintage, override rate, PSI on inputs, latency p99, error rates — the signal risk leadership actually watches.
Alerting on drift
PSI / CSI alerts at input and score level. Calibration drift alerts. Vintage rollout alerts. Quiet until something matters.
Disaster recovery
Active-passive multi-region failover and replay-from-log recovery. Designed to be the layer that stays up when others don't.
What's deliberately absent
Not a generic rules engine. Not a BPM workflow tool. Not a streaming platform pretending to be a decisioning fabric.
FlowX doesn't try to be everything. It is the decisioning fabric for credit and fraud — purpose-built for those two domains, with the bureau orchestration, parsing, attribute computation, A/B testing, and audit primitives those domains demand. If you need a generic workflow engine, there are good ones. FlowX is not one of them.
How it's delivered
With a forward-deployed decisioning team — not as a self-serve product.
Decisioning systems that touch every dollar of origination, every payment authorization, and every claim aren't the kind of system you self-install. Every FlowX engagement ships with a forward-deployed team — decisioning engineers, strategy analysts, and MRM-aligned operators who stand up the fabric, run the strategies, and stay accountable for the outcomes.
Managed end-to-end
We stand up FlowX to your security and compliance posture, and we operate it. You set the controls. We carry the work.
Decisioning practitioners
Engineers and strategy analysts who have run real-time decisioning at issuer, bank, and fintech scale. Not generalist platform consultants.
Accountable end-to-end
From the first integration to the post-launch monitoring review. The same people. Continuity is the point.
In production
FlowX powers origination, line, pricing, and collections decisions on $2.5B+ in annual originations across fintech lenders — from BNPL and installment to personal loans and revolving credit. Every decision logged. Every strategy versioned. Every score defensible.