Skip to content
Mardi Gras Developers
Menu+

Fintech · 2026 – present

A private AI platform for a financial services firm

A financial services firm wanted its staff to work with AI, without customer data leaving Europe and without depending on a single provider. We built its own AI platform with a model gateway, EU routing that blocks when it cannot be guaranteed, and a masked copy of production data that AI can answer questions about.

Updated

7, each with an EU fallback
Model tiers
9, all on pinned versions
Services in the stack
6 SQL Server databases
Source databases
European endpoints only
Processing

Client European financial services firm (name confidential)

The challenge

  • Staff needed AI support, but public tools are not an option for financial and personal data.
  • The organisation had to be able to show where prompts and completions are processed.
  • Analysts wanted to ask questions about production data, which is full of personal data and lives in SQL Server.
EmployeeSSOZero-trustsign-in requiredPII maskingPresidioModel-gateway7 tiers · fail closedEU modelprimaryEU fallbacksecond providerOutside the EUblocked

What we built

  1. 01

    Model gateway with tiers per task

    One gateway with seven model tiers: from fast and cheap for everyday questions to heavy models for deep analysis. Every tier has a fallback route to a second European provider.

  2. 02

    EU routing that fails closed

    All requests are pinned to European endpoints. If a model has no EU endpoint, the EU fallback route takes over; if there is none, the request is refused. Open-source models only run on European infrastructure and are excluded from credit decisions.

  3. 03

    Masked analytics copy

    A repeatable pipeline turns backups of six SQL Server databases into PostgreSQL: restore into a throwaway environment, load, mask, verify and publish. Personal data is replaced deterministically; keys and amounts stay intact.

  4. 04

    Verification as a hard gate

    Before publishing, the pipeline checks row counts, exact equality of money amounts between source and target, the presence of relations and possible PII leaks.

  5. 05

    Search quality measured, not assumed

    An evaluation script measures recall and MRR of search in the knowledge bases. On that basis we chose a deterministic pre-processing step that gives every passage its document title and headings, instead of a costlier variant in which a model rewrites each passage. It needs no model calls and no text leaves the environment.

  6. 06

    A separate, locked-down analyst environment

    Questions to the database go through a dedicated terminal with a role that may only read the masked copy. We added that separation after seeing that a shared terminal could reveal secrets of other services.

  7. 07

    Hardened infrastructure

    Users get in through Cloudflare Zero Trust with single sign-on. The nine services publish no ports to the internet themselves, all images are pinned and containers cannot gain extra privileges. An audit of the running server recorded what was still open and the order in which it can be closed safely.

Outcome

  • Staff work with AI in their own environment, without prompts or completions being processed outside Europe.
  • Analysts can ask questions in plain language about a read-only copy that contains no real personal data.
  • Every architecture choice is recorded with its reason in a decision document compliance can read.
  • The organisation is not tied to one AI provider: switching a model is a configuration change.

Technology

  • Open WebUI
  • LiteLLM
  • OpenRouter EU
  • Amazon Bedrock EU
  • PostgreSQL + pgvector
  • SQL Server
  • pgloader
  • Presidio
  • Docling
  • Redis
  • Cloudflare Zero Trust
  • Docker Compose

Services involved

Contact

Book a 30-minute intro call

Describe your question in a few sentences. Kevin or Pascal replies personally.