CASE STUDIES

Case studies from our work.

Each case study covers one AI system: what our audit or build found, and what our engineers changed.

Browse case studies ↓
CASE STUDIES · HOW WE DESIGN AND BUILD READ THE CASE STUDIES ↓
CASE STUDY 01

Credit decision support: what the audit found, and how we fixed it

The challenge: the lender's credit-scoring pipeline could not show how the model was built, tested or overseen, or meet the duties that apply to creditworthiness systems.

What we built: Feature ingestion on Azure Data, bias and drift checks before each release (Article 10), logs kept for at least six months where the Act applies (Article 19 for providers; Article 26(6) for deployers), human review of declined applications (Article 14), and a combined Article 27 assessment and GDPR DPIA where both apply. Each release now produces a technical documentation pack (Annex IV) for review.

How the scoring, human review, logs and impact assessment fit together, with diagrams.

Annex IV
Technical file for each release
Override
Human-override console
Lineage
Lineage and quality gates
Read the full case study →
ARCHITECTURE · FINANCE Controls by design
DATA LINEAGE · Azure Delta LakeART. 10 CHECKS
BIAS & DRIFT GATE · Python Eval SuiteBEFORE RELEASE
HUMAN OVERRIDE · Underwriter ConsoleART. 14 DESIGN
ANNEX IV & ISO/IEC 42001 EXPORTPER RELEASE
STACK: AZURE DATA · PYTHON · ISO/IEC 42001 · EU AI ACT
CASE STUDY 02

Policy tracking with AI: a working prototype

An AI prototype that reads insurers' drug rules and flags what changed, with people in control.

The challenge: analysts read every insurer policy document by hand. Confirming a single change takes 3 to 6 months, and two analysts can read the same rule differently.

What it does: an AI assistant does the reading and flags what changed, so analysts can focus on reviewing what is flagged and deciding what to act on.

Status
Working prototype
Input
Policy documents
Control
Human review
Read the full case study →
PROTOTYPE FLOW · POLICY TRACKING People in control
INPUT · Insurer drug-policy documentsRECEIVED
READ · AI assistant reads each documentAUTOMATED
FLAG · What changed, surfaced for reviewFLAGGED
DECIDE · Analyst reviews and actsHUMAN DECISION
CASE STUDY 03

Inside an AI policy intelligence platform

The architecture behind a working prototype: managed AWS services, AI extraction and two human review gates.

What it is: a prototype that runs end to end on AWS. AWS Step Functions drives every document through four layers of managed services, with no servers for us to run. Amazon Bedrock handles extraction and classification, and a confidence score routes each change to a path.

Human control: Gate 1 is analyst review and Gate 2 is lead approval, so people keep control before anything is published.

Cloud
Managed AWS services
Review
Two human gates
Audit
Every step logged
Read the technical case study →
PLATFORM LAYERS · AWS Two review gates
ORCHESTRATE · AWS Step FunctionsEVERY DOCUMENT
EXTRACT · Amazon Bedrock Data AutomationAI
CLASSIFY · Bedrock Knowledge BasesTAXONOMY
GATE 1 · Analyst reviewHUMAN
GATE 2 · Lead approvalHUMAN
STACK: BEDROCK · AGENTCORE · STRANDS AGENTS SDK · LAMBDA
CASE STUDY 04

An applicant tracking system: what the audit found, and how we fixed it

An EU AI Act audit of an applicant tracking system for hiring: what it found, and how we fixed it.

The scope: systems that screen or rank job applicants are high-risk under Annex III, point 4 of the EU AI Act. The audit read the screening pipeline, guardrails, storage and interface from source, along with the documents, and ran the system in a test environment.

28
Obligations checked
18
Applicable
High-risk
Tier
Read the full case study →
AUDIT SCOPE · EU AI ACT Audit complete
SCOPE · Annex III, point 4: employmentHIGH-RISK
ROLE · ProviderASSESSED
STATIC SCAN · AI bill of materialsSIGNED
OBLIGATIONS · Obligation map28 MAPPED
RUNTIME · Test data onlyTEST ENVIRONMENT
STACK: PYTHON · POSTGRES OR SQLITE · DOCKER
CASE STUDY 05

Healthcare AI toolkit: what the audit found, and how we fixed it

An EU AI Act audit of a healthcare AI toolkit: what the audit found, and how we fixed it.

  • 5 AI components detected: 3 SDK components and 2 prompt templates
  • 5 findings, all medium
  • 28 provider obligations checked: evidence in full for 1, in part for 12
5
AI components
5
Findings, all medium
28
Obligations checked
Read the full case study →
AUDIT SUMMARY · EU AI ACT Audit complete
TIER · High-risk, set by intended useASSESSED
COMPONENTS · SDKs and prompt templates5 FOUND
OBLIGATIONS · Provider register28 CHECKED
EVIDENCE · Full / partial / none1 / 12 / 4
RUNTIME · Items needing a running system7 FLAGGED
STACK: LANGCHAIN · HUGGING FACE TRANSFORMERS · PROMPT TEMPLATES

Want to talk through a build like these?

Tell us about your workflow, data and any regulatory date. We will set out the controls and build steps we would expect, and set out a first step.

Start a project