Xunodus
All outcomes
Head of Engineering · Regulated fintech

Private AI platform on the client's own servers

A regulated fintech wanted AI in its operations but could not send customer data to third-party SaaS. We built a private platform on its own servers — agents, workflows and an operations console — and handed over the keys.

Representative engagement. This case combines patterns from comparable engagements. Client names, figures and interfaces are illustrative.

Sector
Regulated fintech
Timeline
3 weeks to production
Scope
Private model gateway, workflow engine, monitoring, access control
Stack
Docker · Cloudflare Zero Trust · PostgreSQL · n8n
Illustrative interfaceThe operations console: agent throughput, infrastructure health and executive KPIs, on servers the client owns.
01

The situation

Compliance had blocked every AI pilot for a year. Each tool the operations team wanted — a support agent, document checks, reporting — meant sending customer data to a vendor in another jurisdiction.

So the work stayed manual. Analysts copied data between systems, and the few automations that existed ran on personal laptops with nobody watching them.

02

Where it broke

  1. 01Customer data could not leave approved infrastructure, and most AI tools required exactly that.
  2. 02Automations had no owner, no monitoring and no audit trail.
  3. 03Every vendor added a bill, a login and another place where data could leak.
Illustrative interfaceAgents and workflows in one view: requests flow from WhatsApp through the model layer to CRM, ERP and a human queue.
03

What we built

A private platform on the client's own servers. Agents reach models through a gateway inside the network, workflows run in hardened containers, and every service sits behind Cloudflare Zero Trust with single sign-on.

An operations console shows health, throughput, latency and cost in one place, and every automated action is written to an audit log the compliance team can query. Nothing runs that nobody can see.

System map
  1. Requests
    WhatsApp, internal tools and APIs
  2. Model gateway
    Models run inside the client's network
  3. Workflow engine
    n8n and queues in hardened containers
  4. Ops console
    Health, cost and an audit trail in one view
04

How we rolled it out

  1. Days 1–4
    Architecture and threat model agreed with security and compliance before any code was written.
  2. Days 5–12
    Platform built: container cluster, model gateway, workflow engine, access control and logging.
  3. Days 13–17
    First three workflows moved off laptops and SaaS tools, tested against real cases.
  4. Days 18–21
    Penetration test, runbooks and hand-over; the internal team now ships the next workflows itself.
Illustrative interfaceThe agent layer: models run where the data lives.
05

What changed

100%
Customer data kept in-house
99.99%
Platform uptime
−58%
Third-party SaaS spend

AI pilots that had been blocked for a year went live in a month. The operations team ships new automations weekly, compliance can audit every action, and three SaaS subscriptions were cancelled.

For the team

AI work that compliance had blocked for a year now ships every week — and customer data never leaves.

Representative outcomes from comparable engagements.

06

What we'd tell another team

In a regulated business, the fastest path to AI is not a better vendor — it is infrastructure you control. Once data never leaves, every other conversation gets easier.

Inside the system

Facing the same bottleneck?

Tell us how it works today. You'll get a written architecture sketch — useful even if we never work together.

Message on WhatsApp
Next case study
Head of Sales · Premium retail
WhatsApp sales agent for a premium retailer
Read the case study