Aelix CreditFlow · Lending and credit AI
Credit decisions you can defend, in milliseconds, not days.
Aelix CreditFlow pairs a calibrated, explainable risk model with a complete origination spine. Digital onboarding through Maker-Checker disbursement. So your team underwrites faster, justifies every outcome, and stays compliant across the US, UK, and India.
Aelix Agents · Agentic AI · Now in active build
An agentic layer over the origination spine.
CreditFlow now includes an agentic layer. Autonomous, tool-using agents that act through CreditFlow's governed APIs, do the routine origination work, and escalate the judgment calls to a human. Every agent is human-in-the-loop by default, runs inside your perimeter, and writes each step to the same hash-chained audit log as the rest of the platform.
Underwriting Copilot.
Pulls the application, runs scoring and the rules engine, checks documents, and drafts a decision with reason codes for the credit officer to approve.
Document Chase Agent.
Detects missing stipulations and runs the borrower follow-up loop until the file is complete.
Portfolio Watch Agent.
Monitors concentration and breach signals and opens a remediation task when a limit drifts.
Agents act only through permissioned, least-privilege APIs. They never finalize a lending decision or release funds without a credentialed human approval, and the Maker-Checker gate still stands. The reasoning model is a swappable component: run a local model on-prem for full data residency, or a hosted model where policy allows.
Now in active build. The first agents ship as configurable, versioned workflows.
The problem
Fast or defensible. Most lending stacks force a trade-off no risk officer should have to make.
Most lending stacks force a trade-off no risk officer should have to make. Decisions are either fast or defensible. Rarely both.
Slow.
Applications crawl through disconnected intake forms, manual document chases, and spreadsheet-driven underwriting. Days pass before a borrower hears anything.
Opaque.
'AI' scores arrive as a number with no reason behind it. When a regulator, an auditor, or a declined applicant asks why, nobody can answer with confidence.
Non-compliant by accident.
Adverse-action notices, consent receipts, and disclosure requirements differ by region and product. Bolting them on after the fact is how good lenders end up with bad audits.
CreditFlow is built to close all three gaps at once. With real, working software, not a slide.
Available today
Shipped, running code. Not a roadmap drawing.
Explainable, calibrated risk scoring, not a black box
At the core of CreditFlow is a real, trained, probability-calibrated credit model. A gradient-boosted classifier with sigmoid (Platt) calibration that turns raw model output into a usable 12-month probability of default, then maps it to a familiar 300 to 900 display score.
Every score ships with model-derived reason codes. Each feature is measured against the population baseline to surface its marginal contribution to this specific decision. That means your underwriters, your auditors, and your adverse-action notices all draw from the same explainable source of truth, not a hand-coded lookup table.
On a held-out evaluation cohort, the model reaches an ROC AUC of approximately 0.81 with a Brier score of approximately 0.09. Characterizing a genuine, calibrated learning pipeline rather than a placeholder.
Decisions measured in milliseconds
Speed is not a marketing target. It is a benchmarked number. The CreditFlow decision hot path (inference plus rules evaluation) is measured at approximately 18ms p50 and 21ms p95, comfortably inside a sub-200ms budget. The benchmark is reproducible and writes a JSON artifact your team can wire into CI, so the latency you quote in an RFP is the latency you can prove.
A versioned, region- and product-scoped rules engine
A JSON-DSL rules engine sits alongside the model. Rules are versioned and scoped by region and product, so a Commercial Real Estate policy in the US can differ from a working-capital policy in India without code changes. The engine can override the model to manual_review or auto_decline. The model informs the decision. Your policy governs it.
Role-gated decisions and Maker-Checker disbursement
Underwriting authority is enforced through role-based access control, and every disbursement passes through a Maker-Checker gate before funds move. A built-in separation of duties that also satisfies Strong Customer Authentication under PSD2. Releases run across multiple region-pinned payment rails (FedWire/ACH in the US, FPS/BACS in the UK, NEFT/RTGS/IMPS/UPI in India).
Document stipulations with real OCR
Stipulations are resolved with real local OCR (Tesseract) over uploaded documents, with a deterministic fallback when the engine is unavailable. So the document spine works end to end today, behind a single stable interface that is ready to swap in a managed OCR provider later.
A real portfolio optimizer
CreditFlow's portfolio engine is a working optimizer, not hardcoded figures. It performs concentration-aware allocation against sector limits and risk appetite, then projects risk-adjusted yield, RAROC, expected loss, and capital efficiency, and flags concentration breaches where exposure exceeds limit. Stress testing folds the same expected-loss math, so your scenarios stay consistent with your live book.
Tamper-evident audit, by construction
Every consequential action is written to a hash-chained audit log. Each row carries the prior row's hash, so altering any entry breaks the chain. A verification CLI (aelix verify-audit) walks the chain and recomputes every hash, giving you cryptographic, tamper-evident proof that your decision history has not been touched.
How it works
One continuous spine from application to funded.
CreditFlow is a complete loan-origination spine. One continuous flow from application to funded.
01
Digital onboarding.
Borrowers and applications are captured through guided digital intake with KYC.
02
Document stipulations.
Required documents are collected and parsed with real OCR, then cleared against the file.
03
Rules + scoring.
The versioned, region/product-scoped rules engine runs alongside the calibrated risk model to produce a score, reason codes, and a recommended outcome. In milliseconds.
04
Role-gated decision.
Authorized underwriters approve, decline, or escalate to manual review, with adverse-action and disclosure artifacts generated per region.
05
Maker-Checker disbursement.
A second authorized role approves the release. Funds settle across the appropriate region-pinned payment rail.
06
Continuous oversight.
Every step lands in the hash-chained audit log, feeding monitoring, portfolio optimization, and compliance consoles.
The honest metrics
Numbers we can defend, framed with their conditions.
Calibrated ML scoring · held-out AUC approximately 0.81.
A real trained model, not a stand-in.
Decisions measured at p50 approximately 18ms, p95 approximately 21ms.
Benchmarked, well inside a sub-200ms budget.
300 to 900 explainable score.
Monotone in a calibrated probability of default, with model-derived reason codes.
3 regions, day one.
US, UK, and India compliance scaffolding (FCRA, ECOA, GDPR, PSD2, DPDP, RBI).
Hash-chained audit.
Tamper-evident by construction, verifiable from the CLI.
On the roadmap
Real, in-progress directions. Not things we pretend are live today.
We tier our claims honestly. These are real, in-progress directions, not things we pretend are live today.
Live credit-bureau connectivity.
Bureau adapters are credential-pluggable with an HTTP adapter contract already in place (mock is the default profile). Live Experian, Equifax, TransUnion, and CIBIL calls await vendor credentials and certification.
Live payment-rail settlement.
Rail adapters follow the same pluggable pattern. Live FedWire/ACH/FPS/UPI settlement awaits rail membership and sponsor-bank agreements.
Training on production lending history.
Today's model trains on an engineered, representative synthetic cohort behind a stable scoring interface. Swapping in a licensed real bureau/cashflow corpus changes nothing downstream, and is the path to a model validated on real-world discrimination and fairness. A deep-learning model on production-scale history is a roadmap goal, not a present-day fact.
SOC 2 Type II and ISO 27001.
The security posture is already in the code. The certifications require completed third-party audits, which are in progress.
Security and compliance
Engineered for regulated lending from the ground up.
SOC 2-aligned controls.
Role-based access control, multi-factor authentication, and a hash-chained, tamper-evident audit trail across the platform.
GDPR-ready by design.
Consent capture, data-subject-request tooling, and per-region notice handling are built in. Not bolted on.
Per-region compliance scaffolding.
FCRA and ECOA reason-coded declines and adverse-action notices (US), UK GDPR Article 22 automated-decision disclosure, PSD2 Strong Customer Authentication via Maker-Checker, and India's DPDP consent notices plus RBI Key Fact Statement and cooling-off disclosures.
Certifications such as SOC 2 Type II and ISO 27001 are on the roadmap. The underlying controls they audit are already implemented.
See it on your stack
See CreditFlow on your own credit policy.
Bring your products, your regions, and your rules. We will show you a calibrated score with its reasons, a sub-200ms decision, and a Maker-Checker disbursement. End to end, on real working software.
Frequently asked
Questions teams ask in the second meeting.
Is the model explainable, or is it a black box?
Explainable. Every score carries model-derived reason codes computed from each feature's marginal contribution against the population baseline. Not a hand-coded table. The same explanation drives underwriter review and adverse-action notices. These are directionally informative local explanations. A full SHAP-grade adverse-action artifact is part of our productionization path on real data.
What is the model actually trained on?
Today it trains on an engineered, representative synthetic cohort with a non-linear, interaction-bearing default process. Designed so the model genuinely learns and calibration is meaningful, rather than memorizing. It reaches approximately 0.81 held-out AUC on that cohort. Those numbers characterize the pipeline, not real-world discrimination. The scoring interface is swap-ready for a licensed real bureau/cashflow corpus, which is the prerequisite before any production lending use.
How fast are decisions, and is that benchmarked?
The decision hot path (inference plus rules evaluation) is measured at roughly 18ms p50 and 21ms p95, well inside a sub-200ms budget. It is a reproducible benchmark that emits a JSON artifact, not a target in a slide deck. Database round-trips are monitored separately.
Can I enforce my own credit policy, and keep a human in the loop?
Yes. The versioned, region- and product-scoped rules engine governs the decision and can override the model to manual review or auto-decline. Disbursement always passes through a role-gated Maker-Checker approval, giving you enforced separation of duties before any funds move.
How do you handle multi-region compliance and auditability?
Compliance scaffolding ships for the US, UK, and India. FCRA/ECOA, GDPR/UK GDPR, PSD2, DPDP, and RBI. Covering reason-coded declines, automated-decision disclosures, consent capture, and data-subject requests. Every action is recorded in a hash-chained audit log that a verification CLI can independently confirm is untampered.
Need this tailored to your environment?
Every Aelix product can be configured, extended, or built bespoke for your industry, data sources, and compliance constraints. Talk to our engineers about what would change.
Configurable workflows
Adapt rules, thresholds, and approval flows to match your operational policies.
Custom data integrations
Connect to your specific ERP, MES, SCADA, CRM, or proprietary systems.
Bespoke modules
Build product extensions tailored to your industry, region, or compliance needs.