AI Governance for Regulated Finance

Your AI models are making decisions. Can you prove they're doing it correctly?

Magpie is the AI governance and observability platform built for regulated financial institutions. Self-hosted. Audit-ready.

DIFC Regulation 10 compliant from day one.

Live · enforcement console
prod-ae-east-1
Decisions · 24h
1,284,902
p50
4.2 ms
  • credit-risk-v4.2·applicant_emirates_id
    4msALLOW
  • fraud-net-v1.4·txn_amount_aed
    5msALLOW
  • kyc-classifier-v3.0·merchant_mcc
    6msREDACT
  • aml-screener-v2.1·customer_segment
    3msALLOW
  • intent-router-v0.9·swift_message_text
    7msBLOCK

Trusted by DIFC-licensed fintechs and financial institutions in the UAE

DIFCADGMCBUAEDFSA
DIFC Reg 10PDPL · Federal Decree-Law 45/2021CBUAE Model Risk Guidance 2026DFSA Examination ReadyISO/IEC 42001SOC 2 Type IINIST AI RMFSelf-hosted · No raw data leavesRFC 3161 timestampingSHA-256 chained auditDIFC Reg 10PDPL · Federal Decree-Law 45/2021CBUAE Model Risk Guidance 2026DFSA Examination ReadyISO/IEC 42001SOC 2 Type IINIST AI RMFSelf-hosted · No raw data leavesRFC 3161 timestampingSHA-256 chained audit
01The problem

Your regulators are not waiting.

DIFC Regulation 10 is in force. CBUAE model risk guidance is binding. The DFSA is actively examining how firms govern their AI systems.

Most financial institutions are deploying AI in credit, fraud, compliance, and customer service — and managing governance in spreadsheets, shared drives, and email threads. That is not a sustainable position when an examiner arrives.

01 / Gap
MODELOWNERRISKSTATUS????????????????????????????????????????UNKNOWN

No model inventory

You cannot produce a list of every AI system in production, the decisions it makes, or who approved it.

02 / Gap

No pre-deployment sign-off

Models went live without a structured risk assessment. There is no documented evidence that someone reviewed the risks before the model touched customers.

03 / Gap

No audit trail

You cannot reconstruct what data a model received, what decision it made, or whether a human reviewed it — for any given customer, on any given date.

Magpie solves all three. On-premises, within your infrastructure, with no customer data leaving your environment.

02Capabilities

Governance across the full AI lifecycle.

Eight modules — built specifically for regulated AI in DIFC and UAE financial services. Not just observability after something goes wrong; control before and during the decision.

01

Model inventory

Every AI model your firm runs — in one place. Credit scoring, fraud, KYC, customer service, compliance. Risk tier, version history, ownership, and approval status. Answer the DFSA in thirty seconds.

02

Pre-deployment risk assessment

Templated assessment workflow every model completes before it goes live. Multi-party signed off, exported as a tamper-evident PDF.

03
<10ms

Real-time enforcement SDK

Lightweight in-process SDK that enforces governance policies in the hot path. PII redaction, confidence routing, residency assertions — under 10ms.

04

LLM-as-judge evaluation

A self-hosted small language model evaluates outputs against your declared policies — at scale, without a single record leaving your environment.

05

Immutable audit log

Cryptographically chained record of every model decision. Hash-anchored, RFC 3161 timestamped, exportable in one click.

06

Human oversight log

Every reviewer, decision, rationale, and time-to-decide. Override rates surface automatically — before your regulator finds them.

07
f(x)+-+

Feature schema & explainability

Plain-language feature views for compliance officers, not data scientists. See which features drove a decision and by how much.

08

Board reporting

One-click board report pack: inventory, assessment status, overrides, incidents, upcoming reviews. The artifact your CRO needs.

03How it works

Self-hosted. Your data never leaves.

Not a SaaS platform that ingests your customer data. One pip install. One docker pull. Full governance in production within a day.

01Inference call02SDK · enforce03Policy + LLM judge04Audit chain05Governance UI
Step 01

Your model serving layer

Inference runs where it always has — inside your stack.

Step 02

Magpie SDK

Enforcement agent runs in-process. Wraps inference calls, enforces policy in the hot path.

Step 03

Magpie local agent

Single Docker container in your environment. Stores events in your database with object-lock enabled.

Step 04

Your database & object storage

Raw decision data, feature vectors, audit chain — all in your infrastructure, permanently.

Step 05

Magpie governance UI

Compliance team works on metadata, aggregates, and schemas only. No raw data ever traverses.

04The enforcement difference

Most platforms tell you what happened. Magpie stops the wrong things from happening.

Passive observability is necessary. It is not sufficient. Knowing that a model sent a customer's Emirates ID to an external API is useful — knowing it before it happens is better.

Magpie's enforcement agent evaluates every input before your model sees it and every output before your application acts on it. PII detection, confidence thresholds, output consistency, volume anomaly, residency — in real time, sub-10ms, in your infrastructure.

Inbound prompt
{
  "model": "kyc-classifier-v3.0",
  "input": {
    "emirates_id": "784-1991-3340721-2",
    "iban": "AE070331…"
  }
}
policy gate
Forwarded to model
redacted
{
  "model": "kyc-classifier-v3.0",
  "input": {
    "emirates_id": "[PII:EID]",
    "iban":         "[PII:IBAN]"
  },
  "_magpie": {
    "policy":  "uae_pii_v2",
    "logged":  "sha256:9f4a…b21c",
    "latency": "4.2ms"
  }
}

You define the policies. Magpie enforces them. Every action is logged, timestamped, and auditable.

05Immutability

Your audit log is only useful if it can be trusted.

Magpie's audit log is cryptographically chained. Every event is hashed with the hash of the event before it. If any record is modified or deleted — by anyone, including your infrastructure team — verification fails and your compliance officer is alerted.

When a DFSA examiner asks you to produce decision history for a specific customer on a specific date, you produce it. And you can prove it has not been touched.

Block 001
9f4a…b21c
decision · 12:04:11
verified
Block 002
c318…2af0
override · 12:04:22
verified
Block 003
71de…0a93
policy_hit · 12:04:30
verified
Block 004
55b1…f72e
decision · 12:04:41
verified
Block 005
e0c7…1ad5
decision · 12:04:55
verified
Mechanism
SHA-256
Per-record hash, visible in the UI
Mechanism
RFC 3161
Timestamp authority anchoring
Mechanism
Object lock
Retention-locked storage, no deletion
Mechanism
Nightly verify
Chain re-verified every 24 hours
06Who it is for

Built specifically for regulated AI in DIFC and UAE financial services.

Right customer
  • DIFC-licensed fintech or financial institution with 50–500 employees
  • Running AI in credit, fraud, KYC, compliance or customer service
  • Head of Compliance or CRO responsible for model risk governance
  • DFSA examination coming or a CBUAE model risk review due
  • Need to demonstrate DIFC Regulation 10 compliance
Not for you
  • Not in a regulated industry
  • Looking for a generic LLM evaluation tool
  • Comfortable with customer data leaving your infrastructure
Why not a generic tool

LangSmith, Braintrust and Arize are built for AI teams. Magpie is built for regulated enterprises.

Cloud-hosted

Your credit and fraud data cannot go there. Magpie runs inside your infrastructure.

No pre-deployment sign-off

Generic tools have no concept of templated workflows, multi-party attestation, or regulatory audit export.

No regulatory awareness

They do not know what DIFC Regulation 10 requires. They do not know what a UAE IBAN looks like.

07Compliance coverage

What Magpie maps to — specifically.

DIFC Regulation 10 — Autonomous Systems

  • Model inventory
  • Pre-deployment risk assessments
  • Human oversight configuration
  • Explainability for consequential decisions
  • Audit trail

PDPL — Federal Decree-Law 45/2021

  • PII detection and redaction at the enforcement layer
  • Data residency controls
  • Purpose limitation enforcement
  • Consent-aware processing flags

CBUAE AI/ML Model Risk Guidance (2026)

  • Board-level reporting pack
  • Model validation workflows
  • Override rate monitoring
  • Incident management
  • Change control

DFSA Examination Readiness

  • One-click audit export
  • Assessment PDF artifacts
  • Immutable decision log
  • Enforcement event log
  • Schema version history
08Pricing

Straightforward. No per-seat surprises.

All plans are self-hosted. Annual contracts. Implementation support included. Your data never leaves your infrastructure.

Tier

Pilot

For teams validating Magpie against a single model in a staging environment.

  • Up to 2 models
  • Up to 100,000 decisions/month
  • Full feature access
  • 90-day term
Most chosen

Growth

For firms with multiple production models and active compliance obligations.

  • Up to 10 models
  • Up to 2 million decisions/month
  • Board reporting
  • SLA
Tier

Enterprise

For institutions with complex model estates, custom policy or multi-entity deployments.

  • Unlimited models
  • Custom decision volume
  • Custom SLM-as-judge deployment
  • Dedicated support
09Get started

Your next DFSA examination is not a date to prepare for.It is a date to be ready for.

Magpie can be deployed in your staging environment within a day. Your first model inventory takes an hour. Your first risk assessment takes an afternoon. The firms that will be ready are the ones who start before the examination is announced.

Magpie.

AI Governance and Observability for Regulated Financial Institutions.

CoverageDIFC Regulation 10PDPLCBUAE Model Risk Guidance
Built forDIFCADGMUAE financial services
DeploymentSelf-hostedNo raw data leavesAudit-ready
© 2026 Magpie. Product of Steinn Labs.Based in Dubai, UAE