Client: a European energy company (trading and dispatch)
Engagement: 2022 to today
Role: started as full-stack engineer, now accountable for bringing a business-critical system into production

Most case studies describe a system. This one describes a job that changed three times, and why the last version of it is the one that matters for the next few years of software.

Phase 1: Building new trading tools

The trading desk needed a single view of deals, positions and valuations instead of spreadsheets spread across several systems. I joined when there was little more than an idea and helped build the platform from the first commit: data model, backend services, trader-facing UI, cloud infrastructure and the pipelines to run it.

Stack: TypeScript, Angular, Python, AWS serverless, a cloud data warehouse.

Phase 2: Modernizing a legacy core system

A long-lived Java system at the heart of gas operations had to start consuming data from new event-driven platforms. That meant first bringing it up to a current Java platform, then designing a reusable Kafka integration layer with schema management and reliable, exactly-once style processing, validated through long-running tests before it went live.

Stack: Java, Kafka, Avro, CI/CD.

Phase 3: Accountable for the path to production

The latest phase moved that system onto new infrastructure. My role shifted from implementing features to owning the whole path to production:

  • Deployment automation, containerization and monitoring
  • Architecture decisions, documentation and review boards
  • Security and quality gates, risk assessments
  • Cutover planning, runbooks and handover to operations

Bringing AI into a regulated environment

During this phase the company opened governed access to large language models through its own cloud environment and actively encouraged teams, including the business side, to use them. I used them daily in engineering work and saw first hand what makes an enterprise rollout succeed:

  • Governed access beats shadow AI. Once models ran inside existing identity, data and compliance boundaries, the question changed from “are we allowed to?” to “what can we do with it?”
  • Context is the product. The same model gives very different results depending on what you feed it. Clear specs, architecture docs and domain knowledge were the real multiplier.
  • The biggest gains weren’t only in code. Documentation, analysis, test design and review preparation sped up at least as much as implementation.
  • Accountability stays human. AI speeds up the work, but someone still has to own the spec, the review and the decision to go to production.

What changed

In 2022 the hard part was writing the code. By now the hard part is everything around it: getting a precise spec, passing security and architecture review, proving a system can be operated, and handing it over cleanly.

AI coding agents are accelerating that shift. Implementation keeps getting cheaper. What stays expensive is knowing exactly what to build, and being accountable for getting it safely into production in environments where governance is the real bottleneck.

That’s how I work now: AI agents for the heavy lifting, my own time on specs, context, review and the path to production.

Working on something similar? Get in touch: [email protected]