Colin Olliver

Killing the Daily Failure

When I joined, the Databricks platform was failing every 24 hours. This wasn't a minor inconvenience — it was a critical data platform serving a €100m global messaging business processing billions of messages annually. The vendor who built it couldn't diagnose the root cause.

Through systematic investigation, I identified the failures stemmed from multiple compounding issues: poor pipeline architecture, cluster configurations mismatched to workload types, no fault tolerance, and no pre-processing data storage layer.

What I did:

• Optimised pipeline architecture from the ground up
• Implemented workload-tuned cluster configurations matched to actual usage patterns
• Built fault-resilient pipelines with automatic retry mechanisms
• Introduced pre-processing data storage layers

The result:

Eliminated all failures completely. The platform has maintained zero downtime for 18+ months across multi-brand, multi-region operations handling millions of daily events.

Beyond stabilisation, I re-architected the wider Databricks estate — realising €500k in annual savings while simultaneously doubling platform capacity.

Red-Teaming Microsoft's AI

Before deploying Azure AI services to customers across our 20+ messaging brands, I led comprehensive red-team security testing. This wasn't a box-ticking exercise — we were about to expose AI features to real customers across multiple jurisdictions with different regulatory requirements.

I needed to know the content safety guardrails actually worked. They didn't.

What I found:

Vulnerabilities in the content safety filters that could have created regulatory and legal exposure across multiple jurisdictions. The guardrails that were supposed to prevent harmful outputs had exploitable gaps.

What I did:

• Designed and executed systematic red-team testing protocols
• Documented vulnerabilities with clear reproduction steps
• Escalated findings directly to Microsoft's AI safety team
• Collaborated with Microsoft to remediate the risks before go-live

The result:

Safeguarded a platform serving 20+ messaging brands globally. The vulnerabilities were remediated through direct collaboration with Microsoft before any customer exposure. This work established AI safety testing as a non-negotiable step in our deployment process.

From AI Sceptics to 4x Adoption

When we introduced AI tools, the reaction wasn't excitement — it was fear. People thought AI would replace them, didn't trust the outputs, or simply didn't know where to start. Generic "intro to AI" workshops weren't going to cut it.

My approach:

• Designed bespoke Google Gemini training for each business function
• Coached teams on prompt engineering with their actual workflows
• Showed real, tangible value using each team's own use cases
• Removed misconceptions and addressed fears directly

Beyond business adoption:

I built the business case that secured C-suite commitment to AI tooling across the organisation — covering safety, LLM integration, and AI-assisted development.

On the engineering side, I rolled out AI-assisted development starting with Cursor AI and progressing to multi-agent orchestration with Claude Code, boosting developer productivity by 33%.

The result:

Quadrupled monthly process invocations across the business. Went from AI scepticism to AI being embedded in daily workflows across every function.

From Copilot to Crew: Single Agents to Multi-Agent Orchestration

We started by rolling out Cursor AI as a single-agent coding copilot across the engineering team. It worked — 33% productivity boost. But single-agent tooling hits a ceiling: one model, one context window, one task at a time. The developer still orchestrates everything manually.

The evolution:

We moved to Claude Code to introduce multi-agent orchestration — where engineers direct multiple AI agents working in parallel across different parts of a codebase or workflow.

Instead of a developer using one AI assistant, they're now orchestrating a crew: one agent handling tests, one refactoring, one writing documentation, one reviewing code.

Key learnings:

• What we measured with single-agent tooling and where it plateaued
• What multi-agent orchestration actually looks like in practice
• How we're managing quality and trust when agents work autonomously
• The cultural shift from "AI helps me code" to "I direct AI systems that deliver"

The shift:

This is early-stage — we're still learning — but the direction is clear. The shift from using AI as a tool to directing AI as a workforce changes the economics of engineering teams fundamentally.