Systems I designed, shipped and hardened

The problem

The work spans three problems that are usually treated separately but fail together: a generative-AI feature that needs real documents behind it, a data platform that has to be trustworthy, and an Azure landing zone that satisfies security without slowing delivery. My job was to own all three — design the AI solution, build the infrastructure it runs on, and keep the data path healthy in production.

Shape of the solution

Decisions that mattered

• Retrieval instead of fine-tuning first. When answers need to cite a source, grounding beats fine-tuning: it is cheaper to update, and it makes wrong answers traceable. Fine-tuning stays in the toolbox for style and format, not for fresh facts.
• Private endpoints and Key Vault everywhere. Storage, Cosmos DB and SQL sit behind private endpoints with managed identity — no long-lived secrets in pipelines or application settings, which is also what makes rotation a non-event.
• Everything reproducible in Terraform. Environments are built from versioned code, so a new environment is a pipeline run, not a checklist. Reviewing a pull request that adds a subnet is the cheapest security review there is.
• Bastion instead of public admin surfaces. Administrative access goes through Azure Bastion, which removes the "temporarily open the port" pattern that quietly becomes permanent.

What made it hard

Production generative AI is mostly an integration problem. The interesting failures are not in the model — they are in the seams: documents that are technically retrievable but useless, prompts that quietly truncate, pipelines that succeed while producing nothing, and cost that surprises you because nobody measured tokens per request.

What I would do differently

• Instrument retrieval quality before prompt tuning — measure whether the right passage is being found at all.
• Version prompts and indexes the way code is versioned, so a regression has a cause you can point at.
• Add an evaluation set early, even a small hand-labelled one.

The problem

Build cloud infrastructure that is repeatable, secure by default, and reviewable — in a cloud (Oracle Cloud Infrastructure) with a different service model from the one I use daily. The exercise is less about clicking and more about encoding decisions.

Shape of the solution

Decisions that mattered

• Compartment isolation over one big compartment. OCI compartments are a real boundary; separating network, compute and registry means a policy mistake in one place is not a compromise in all of them.
• Least privilege as code. IAM policies are Terraform resources under review, not console clicks that nobody remembers six months later.
• Secrets never in the repository. Gitleaks runs on every commit so a leaked key is a failed build, not an incident.
• Infrastructure scanning is part of the pipeline. Checkov findings are treated as defects with owners, not a report nobody opens.

What I learned

Terraform's real value is not brevity — it is that state becomes reviewable. Once infrastructure is a diff, "who changed the security list" stops being archaeology. The second lesson was about drift: anything not in code comes back, and it comes back at the worst time.

The problem

Professors across faculties spend hours preparing lectures, and personalising material for every student is impractical. The session's question was narrow and practical: where does AI genuinely help, and where does it quietly make things worse?

The pipeline I demonstrated

The point that landed with the room: AI is not here to replace educators, it is here to remove the mechanical parts of preparation so the teaching gets more time.

Try the pipeline yourself

The AI Lab has the lexical half of this running in the browser — chunking with overlap, TF-IDF indexing, cosine ranking — plus an embedding explorer showing why semantic search retrieves what keyword search misses.