Credit origination that runs as deterministic code, with the model invoked only on ambiguous documents. Entry into critical environments is always shadow mode — in parallel with the real system until value is proven. On-premise, with open models. Compatible with banking secrecy.
A credit file is folders of documents, manual validation, and re-keying between systems. AI stays in demos because nobody will let a model near the core.
Intake and validation run as deterministic code. The model judges only ambiguous documents, then hands back to the rail. Proven in shadow mode before it touches anything. Every action recorded.
A consumer credit application arrives: payslips, ID documents, bank statements, a signed contract. The old path is a person opening each document, checking it against policy, and re-keying results between systems until, days later, the disbursement clears.
The new path: intake and classification run as code. Eight documents validate deterministically. One — a payslip photographed at an angle — is escalated to the model for judgment, and its answer is validated again before the file moves. From there the flow returns to the deterministic rail: scoring inputs, approvals, disbursement, record.
KYC & onboarding is under way on the same foundation with an institution in Spain. Same pattern: code where it can be, judgment where it must be, everything recorded.
Every workflow in a bank terminates in a core system that nobody wants to touch. Decades of accumulated logic, dependencies nobody fully maps anymore, and a simple rule learned the hard way: every change is a migration project, and every migration project is a board-level risk.
The data can't leave, either. Banking secrecy means customer records do not flow to external model APIs — full stop. That single constraint rules out most of what gets sold as "AI for banking," because most of it is a thin layer over someone else's cloud.
Manual approval gates on everything preserve safety but erase the point. What a bank needs is a way in that doesn't bet the core: proof in parallel, on infrastructure it already controls.
Credit origination runs inside the bank's own platform as code, with the model invoked only where a document is ambiguous. In critical environments the workflow first runs in shadow mode — in parallel with the real system — until the match rate proves it.
The same platform reaches the problem banks have deferred for decades: legacy code renewal. Agents locate technical debt, renew it, and test it — inside the bank's own platform, without migrating the core. Senior engineers stay where they belong: as the final verification on everything that ships.
Delivery accelerates the same way. Development and review agents multiply the team's capacity, and the code they produce is secured by default — the platform enforces it, not the discipline of whoever wrote the prompt.
All of it runs on-premise with open models, compatible with banking secrecy. The model runs where the data lives. The data never leaves.
Credit origination and disbursement. KYC & onboarding — under way with an institution in Spain. Legacy renewal without core migration. Delivery acceleration across the development organization. Four programs, one foundation: deterministic workflows, the model only where judgment is needed, and security as infrastructure underneath all of it. Entry is always the same — a focused pilot in shadow mode, weekly demos, expansion only after the value is proven. The knowledge stays in the bank.
Entry into critical environments is always shadow mode — parallel to the real system until value is proven.
The system that runs the workflow writes the record. Every document, judgment, and approval is reviewable.
The platform works alongside the core, inside the bank's perimeter. Nothing is ripped out to start.