$ cat ./blog
Notes from building backend systems, AI pipelines, and the occasional open source detour.
One CV per job, a hundred career pages, and interviews where you have to explain architecture you built with AI. The design decisions behind the open-source platform I built to track, apply, tailor and prepare.
Five read-only PostgreSQL queries that find the usual causes of a production slowdown: lock chains, idle transactions, connection exhaustion, unindexed foreign keys and integer IDs running out.
Our search endpoints are hitting URL length limits with GET, and POST lies about semantics. The HTTP QUERY method finally offers both a request body and safe, read-only semantics.
The same AI can act like a careful engineer or a creative storyteller — and the difference is one number. What LLM temperature really controls, and what it taught me about machine creativity.
In 2025, more AI usage meant more productive. In 2026, finance has entered the chat. The question is shifting from 'should we use AI?' to 'how do we use it efficiently?'
A $20 AI subscription used to feel unlimited. Now developers burn a month's quota in days. The cloud computing bill history suggests what's actually happening.
Every company has that one developer who knows why everything was built. When critical context lives only in someone's head, every resignation becomes a risk assessment exercise.
PM: 'Why is there a bug?' Developer: 'Let me ask the AI that wrote it.' A joke about 2026 software development — and the genuinely new debugging problem hiding inside it.
Software isn't just code — it's a collection of decisions. As AI writes more of the implementation, preserving the reasoning behind it becomes the most valuable documentation work there is.
Every developer on your team is prompting AI with a different context — and getting individually valid, collectively inconsistent code. Architecture documents just became more valuable, not less.
AI-generated code can look finished while quietly becoming unmaintainable. The real test of software isn't whether it works today — it's whether someone can confidently change it in six months.
Lessons from shipping a production RAG pipeline — chunking strategies, retrieval quality, cost control, and why your eval suite matters more than your vector database.
How we evolved a Django monolith into an event-driven system — outbox patterns, idempotent consumers, and the mistakes worth avoiding.
What I learned publishing scheduler-calendar — API design for strangers, the documentation bar, and why TypeScript support is the whole product.