Led by a founder with more than 21 years of experience

contact@guilda.io

Case study · Insurance

A global insurer turns detection engineering into a governed platform

Detections used to be written, documented, reviewed and tuned by hand. Now every detection follows one standard, passes automated quality gates and tests before production, and is measured once it is live.

Client profile

Sector
Insurance
Size
More than 2,000 employees
Region
the United States

The challenge

Detections built by hand, with no common standard

The insurer's security operations team protects a business that runs worldwide, and it depends on detections to find threats. Each detection was written, documented and reviewed in its own way. Documentation, reviews, quality checks and tuning requests were manual and slow, and nothing checked a detection automatically before it went live. False positives added to the analysts' workload and to alert fatigue. There was no single view of how well the detections worked, so it was hard to know which ones to improve. Auditors and risk teams also needed evidence of how each security policy was put into practice, enforced, tested and kept up to date, and that evidence was hard to gather.

Our approach

How we built it

We built one platform that governs the whole life of a detection, from its design to its performance in production.

  1. One version-controlled home for every detection

    Every detection and its documentation lives in a single repository under version control. It is the authoritative record: each change is traceable to who made it, when and why.

  2. A standard design, with documentation generated for it

    We set common practices for designing detection logic and automated the documentation that goes with it, so every detection is described the same way without engineers writing it by hand.

  3. AI agents in the engineering workflow

    Agentic AI drafts documentation, carries out first reviews of new detections and speeds up routine engineering work. Its output goes through the same quality gates and approval as any other change.

  4. Quality gates and four kinds of test

    A CI/CD pipeline validates, approves and deploys each detection. Golden, negative, smoke and regression tests check the logic, and a detection that fails a quality gate does not reach production.

  5. A structured route for tuning requests

    Requests to improve a detection follow one process, from the request to the fix, through the same pipeline and tests.

  6. Measuring every detection

    The platform tracks true and false positives, true and false negatives, precision, recall and F1 score for each detection, so the team can see which detections work and which need tuning.

What the insurer kept

  • Detection engineering platform

    An authoritative, version-controlled repository for every detection and its documentation.

  • CI/CD pipeline with quality gates

    Automated validation, approval and deployment of every detection to production.

  • Automated testing framework

    Golden, negative, smoke and regression tests that check detection logic before it goes live.

  • AI-assisted documentation and review

    AI agents that draft documentation and review detections inside the engineering workflow.

  • Tuning request process

    A structured way to raise, track and fix requests to improve a detection.

  • Detection performance analytics

    Precision, recall, F1 score and true and false positive and negative rates for every detection.

Results

What changed

Every detection now follows one engineering and documentation standard

Detections are designed and documented the same way across the team.

No detection reaches production without passing automated quality gates and tests

Golden, negative, smoke and regression tests run on every change in the pipeline.

Documentation, reviews and deployment no longer depend on manual work

Automated workflows, CI/CD and AI agents carry the routine steps.

The team measures how well each detection works and improves it with evidence

Standard detection metrics show which detections to tune, and every change is measured again.

False-positive patterns are found and fixed, helping reduce alert fatigue

This helps the insurer's analysts spend less time on alerts that should not have fired.

Auditors and risk teams can trace how each security policy is put into practice and whether it works

Version-controlled records and test results show how each control is implemented, enforced and validated.

The client is anonymised and gave permission to publish this case study.

Scope

What the work covered

  • Detection engineering
  • Detection as code
  • CI/CD
  • Agentic AI
  • Security operations

Related

The services behind this work

Threat Detection & SIEM Implementation

SIEM implementation on Microsoft Sentinel, Splunk, Elastic Security or Wazuh, with log sources, detection rules and runbooks handed to your team.

Software Development

Software development services for web applications, APIs and integrations, and mobile apps, in TypeScript, React, Next.js, Python, Java, .NET or Go.

AI & ML Implementation

AI implementation services: a use-case assessment, then AI assistants, automation and machine learning models, with the data and platform behind them.

Your turn

Want detections you can test, measure and prove?

Tell us how your team builds and tunes detections today, and we will show you how a governed pipeline would fit.