Learnings from my internship at Capgemini/Telia

My internship at Capgemini/Telia gave me something I didn't expect: not just technical growth, but a much sharper understanding of how serious software gets built and shipped in enterprise environments — and where AI fits into that picture. These five experiences stood out.

AI-Enhanced Coding

I've been using AI coding agents since 2023, but this internship was where I learned to direct them rather than follow them. The difference turned out to be clarity — knowing the architecture, the constraints, and the tradeoffs before writing a single prompt. Working with Claude Sonnet/Opus, Codex, and Gemini, I found that the quality of AI output tracks directly with how precisely you can define the problem. That discipline has permanently changed how I approach new work in a responsible way.

Docker & Kubernetes

Docker became my favorite tool of the period — not just for local development, but as a deployment platform. I built Demo Garden from scratch as an exploration of how Google Cloud Run manages containerized applications internally: on-demand serving, scale-to-zero, high availability.

At Telia, I also worked alongside data engineers running complex on-prem Kubernetes clusters for production data pipelines. That exposure gave me a concrete picture of how Kubernetes applies at scale — both in cloud-native and hybrid infrastructure contexts.

Google Cloud Platform & Event-Oriented Architecture

I completed the GCP Professional Cloud Developer study path, focusing on service inter-connectivity, event-driven design, and cloud governance. The governance piece was a important level-up: I came away confident building solutions that handle sensitive data with proper IAM controls and access boundaries — not just functional applications, but responsible ones.

RAG-Powered Chat Assistant with ADK

This was the most complex project of the internship. I helped develop chat assistants for Telia capable of handling a variety of cases, from product information retrieval to real-time support using internal knowledge bases like Confluence and Jira, to support ticket creation.

The technical surface was wide: ADK multi-agent orchestration, RAG pipeline design, similarity search, prompt engineering, and session context management. But what surprised me most was everything around the code — data privacy requirements, hallucination mitigation, token cost modeling, access control, and navigating alignment across engineering, architecture, and legal stakeholders. Enterprise AI is an entirely different discipline than small scale AI.

I also implemented Langfuse for observability, built an evaluation dataset generator (Evalgen), and developed an alternative GCP Logging integration — giving the team meaningful visibility into agentic system behavior.

See some of this work in action here:

MCP-Enabled Application Development

Toward the end of the internship I got hands-on with MCP (Model Context Protocol), building applications that expose tools and context to LLMs in a structured, composable way. This is where the agentic AI space is heading — moving from single-model calls toward systems where models can reliably act, retrieve, and coordinate. Having practical experience here feels like being early to something that's about to matter a lot.

Going forward

I'll be wrapping up this internship with a clearer picture of what I want to build next: data and AI systems that are production-grade, observable, and built with the full stack in mind.