Built a Supply Chain Finance Platform with Privacy-Safe AI Underwriting — Starting with the Client Portal

A non-bank lender pays clients' suppliers upfront and extends the clients' payment terms under trade credit insurance. Clients submitted invoices as emailed spreadsheets, and onboarding a new client took about two weeks of manual underwriting. Off-the-shelf AI tools could not meet two constraints: identifying applicant data could not reach a language model, and every credit decision had to be explainable and reproducible. Itexus de-identified documents before any model call and placed deterministic credit rules ahead of the model. Delivery started with the portals.

Technologies: Python, FastAPI, PostgreSQL, React, Keycloak, Temporal, Docling, Presidio

Industry	FinTech — Supply Chain Finance / Commercial Lending
Segment	Non-bank lender, insurance-backed payables finance for mid-market companies
Geography	North America and Europe
Deployment Model	Cloud, single-tenant deployment on a multi-tenant-capable architecture
Duration	7 months
Team Size	~9 engineers (Delivery Manager, Business Analyst, Solution Architect, Backend Engineers, Frontend Engineers, QA Engineer, DevOps Engineer, UI/UX Designer)
Delivery Model	Phased delivery: fixed-price discovery → phased build of portals → capped T&M for AI underwriting
Services	Lending Software Development · AI/ML Development · FinTech Product Discovery · UI/UX Design
Technologies	see Technology Stack
Compliance	Designed for ECOA / Reg B, SR 26-2, GDPR and the EU AI Act; SOC 2 Type II-ready

The client is a non-bank lender that finances trade payables for mid-market companies. It pays the client's suppliers directly and later bills the client on extended terms. Facilities are unsecured and wrapped by third-party trade credit insurance. The insurer makes its own independent approval decision before any funding.

The platform handles banking details, audited financial statements and credit decisions for applicants in several jurisdictions. That placed it under US fair-lending rules and, for European applicants, under GDPR. The lender also planned to license the software to other funders. Every architectural and licensing choice had to survive redistribution, not just internal use.

Faster AI underwriting that never exposes applicant identities to a model and keeps every decision explainable.

Business Challenge

• Clients had no portal and emailed invoices as spreadsheets, so staff re-keyed every row into a third-party legacy system by hand.

• Underwriters spent about two weeks preparing each new client's memo by hand.

• No funding could go out without the insurer's approval, and a reduced or withdrawn limit could make already-submitted invoices unfundable.

• The lender wanted to own the platform and license it under other brands.

• Declined applicants were owed specific reasons, and any decision had to be reproducible months later under the policy then in force.

Technical Challenge

• Consistent pseudonymization is cross-document entity resolution: abbreviations, legal-form variants, OCR noise.

• Identifiers also sit in logos, signatures and letterheads, beyond text-based NER.

• Multi-page financial tables from scanned PDFs had to be extracted without numeric drift.

• A 5-minute SLA covered OCR, detection, rules evaluation and the LLM call.

• Insurer limit revocations had to recompute capacity for in-flight invoices without race conditions.

Itexus delivered the platform with a ~9-person cross-functional team over 7 months, deployed on AWS. Delivery started with a discovery phase and proofs of concept on the two riskiest components. The client and operations portals shipped next, and AI underwriting followed once its accuracy target could be measured.

Client Portal and Facility Visibility

• Facility overview showing limit, amount outstanding, available capacity and amounts coming due.

• Bulk CSV/Excel invoice upload with row-level validation, suggested corrections and a check-answers step before submission.

• Capacity preview that recalculates available funding as failed rows are corrected or excluded.

• Searchable invoice register with a full status history for every invoice.

• Supplier management with three bank-detail formats: US routing, UK sort code and IBAN/BIC.

• Repayment schedule and account statements; the platform records payments and never moves money.

Lender Operations Desk and Insurer Gating

• Review queue for approving, rejecting with structured reason codes, or requesting more information.

• Insurer decisions stored as dated, versioned records supporting nil limits, reductions and withdrawals.

• Funding blocked automatically until an insurer approval is recorded.

• Facility limit management and portfolio-level exposure views.

• Append-only, hash-chained audit trail covering every decision and configuration change.

De-Identified AI Underwriting

• Public application flow with verified sign-up, e-signed NDA and financial-statement upload.

• Layout-aware extraction for digital PDFs, with an OCR path for scanned statements.

• Layered entity detection: deterministic patterns, NER models, zero-shot role detection and structural rules for signatures and letterheads.

• Entity resolution that maps every mention of one company to one stable token across the document set.

• A post-redaction verification sweep that fails closed and routes any residual hit to human review.

• Deterministic credit-box rules evaluated before the model, with thresholds stored as data rather than code.

• A policy-conditioned language model drafts the credit memo; underwriters decide, and borderline cases route to a human rather than being declined automatically.

• Controlled re-identification for authorized staff, with every reveal written to the audit trail.

Platform Foundation for Future Scale

• Multi-tenant-capable data model and identity realms, so the lender can run partner programs or license the platform.

• Tokenized design system that re-brands the portals per tenant without code changes.

• Documented REST API, so planned intake channels (API, SFTP, ERP connectors) plug into the same validation pipeline.

• Licence gate in CI that fails the build on copyleft or non-commercial dependencies.

• Decision log that tags each outcome with its policy version, ready for future model validation.

Use Case	Description	Output
Invoice batch submission	Client CFO uploads a batch and fixes failed rows before submission	Validated batch with ID and capacity impact
Invoice review	Lender staff approve, reject or query submitted invoices	Decision record with structured reason code
Insurer gating	Staff record the insurer's decision for a facility	Versioned insurer record; funding unlocked or blocked
Repayment tracking	Client sees what was funded and when it is due	Repayment schedule and account statement
Applicant pre-screening	Prospect applies and uploads financial statements	Indicative, non-binding pre-approval
Credit memo drafting	Model drafts a memo from de-identified statements	Draft memo tagged with policy version
Controlled re-identification	Authorized underwriter reveals the real applicant	Audit entry with user, time and reason

• Credit memo preparation cut from about two weeks to 4 hours.

• Invoice intake moved out of the inbox. Clients submit and correct batches themselves, so staff no longer re-key spreadsheets.

• Insurer risk became visible before funding, not after. Staff see which invoices become unfundable if an insurer cuts a limit.

• AI assistance without identity exposure. The model sees only de-identified content, and every step is logged and reproducible.

• An owned, licensable asset instead of a rented system. The lender owns the code and every dependency permits sublicensing — the first product on a platform, not a one-off tool.

The platform separates identity from analysis: identifying data lives in a segregated vault that the model-calling services cannot reach.

Layer	Components	Function
Experience	Client portal, operations desk, public application flow (React)	User interaction, validation feedback, reporting
Application & Workflow	FastAPI services, invoice state machine, durable workflow engine	Invoice lifecycle, insurer gating, async underwriting pipeline
Decisioning	Rules engine with decision tables, policy versioning, LLM adapter	Credit-box evaluation first, memo drafting second
De-Identification	Extraction, layered entity detection, entity resolution, verification sweep	Removes identity before any model call
Data & Audit	PostgreSQL, segregated identity vault, append-only hash-chained log	Records, token mapping, tamper-evident history
Security	Keycloak (MFA, RBAC), KMS, field-level encryption, zero-retention model endpoint	Access control, encryption, provider data isolation

• E-signature — embedded click-to-sign NDA in the application flow; a simple electronic signature is legally sufficient (Dropbox Sign)

• LLM service — credit memo drafting on de-identified content only, under zero-retention and no-training terms (Azure OpenAI)

• Email delivery — account verification and recovery, submission and decision notifications

• Insurer — no integration by design; the insurer decides independently and staff record its decisions in the operations desk

• ERP and file channels — planned: API, SFTP and ERP/accounting connectors (e.g. NetSuite) feeding the same validation pipeline; CSV/Excel upload at launch

• KYC/AML — planned; the provider is chosen once launch compliance requirements are confirmed

• SSO — planned; email and password with MFA at launch

• Internal ledger and funding instruction export

Category	Technologies
AI/ML	Presidio, spaCy, GLiNER, Azure OpenAI
Document Processing	Docling, pdfplumber, Tesseract / OCRmyPDF
Backend	Python, FastAPI, SQLAlchemy
Decisioning & Workflow	GoRules ZEN, Temporal, XState
Frontend	React, TypeScript
Data	PostgreSQL
Infrastructure & Security	AWS (EKS, RDS, KMS, WAF), Keycloak, Terraform, CI/CD

Security was treated as the first design constraint. The architecture had to prove, not just claim, that identity never reaches the model.

Security Measures Implemented

• MFA for all users and role-based access control on every endpoint.

• Field-level encryption for banking details and the identity vault.

• Identity vault on separate storage, credentials and keys, with no network path from model-calling services.

• Model endpoint configured for zero data retention and no training, confirmed in writing by the provider.

• Append-only audit trail with hash chaining and attributed re-identification events.

• Dependency and licence scanning in CI; separate dev, staging and production environments.

Compliance Considerations

• Fair lending (US): deterministic rules produce specific principal reasons for adverse decisions, as 12 CFR 1002.9(b)(2) requires.

• Model risk: credit-box rules stay deterministic, and deterministic rule-based processes fall outside the model definition in SR 26-2.

• GDPR: the design is treated as pseudonymization, not anonymization, with a documented residual-risk review.

• EU AI Act: corporate credit assessment falls outside Annex III 5(b), and any natural-person path (e.g. guarantors) is gated separately.

• SOC 2 Type II: the architecture is designed to generate audit evidence continuously. This is readiness for a future audit, not certification.

Delivery Model: Phased delivery — fixed-price discovery, fixed price per phase for the portals, capped T&M for AI underwriting until the accuracy threshold was agreed.

Team Composition

Delivery Manager · Business Analyst · Solution Architect · Backend Engineers · Frontend Engineers · QA Engineer · DevOps Engineer · UI/UX Designer

Governance

• Acceptance criteria agreed for each feature before development.

• Proofs of concept on the two riskiest components before committing to a build price.

• AI underwriting accepted against measured entity-detection accuracy on a client-supplied test set.

• Fixed review windows for milestone acceptance built into the schedule.

• Licence gate on every build to protect the client's right to sublicense.

• Weekly progress and budget reports; scope changes approved in writing.

The lender prepares a credit memo in 4 hours instead of about two weeks, on a platform it owns and can license.

Itexus builds lending platforms in which regulatory constraints shape the architecture from the first sprint.

Lending Software Development

Facility, invoice and repayment logic built for a lender that records money flows but never moves funds.

AI/ML Development

A de-identification pipeline and policy-conditioned memo drafting, where the model assists underwriters and deterministic rules own the decision.

UI/UX Design

An exception-first client portal designed to replace spreadsheet workflows for finance teams.