By Oguzhan Ozturk
Senior Data Engineer, Garanti BBVA
A 3-part series
What turns a capable model into an agent you can rely on: the loop, the tools, memory, boundaries and proof, built up around one real task.
Part 1
The Model Is Not the Agent9 min read
Part 2
Memory, Boundaries, and Proof11 min read
Part 3
One Use Case, End to End17 min read
, 4 min read
A map of the eleven open-source repos behind the way I work with coding agents now — grouped into the three things they actually fix: how the agent works, what it knows, and how much of the loop you have to crank yourself.
, 5 min read
Most people use Claude Code like a chat box — prompt, wait, copy, fix, repeat. The next step is to design the loop that prompts the agent for you. This is the install-and-use walkthrough: scaffolding a loop, feeding it fresh signal, and the heavier autonomous engine behind it.
, 5 min read
Claude Code fails in two quiet, expensive ways: it forgets what it learned last session, and it re-reads your whole codebase to answer a small question. This is the install-and-use walkthrough for four tools that fix both — with exact commands and the order I'd add them.
, 6 min read
A skill is a reusable instruction file that teaches Claude Code how to do a task the way a senior engineer would. This is the install-and-use walkthrough for four public libraries — exact commands, when to reach for each, and how to write your own.
, 3 min read
A practical CDC pipeline: stream inserts, updates, and deletes from MySQL into ClickHouse in under five minutes — using Debezium, Kafka, and ReplacingMergeTree to keep an append-only columnar store correct.
Long-form case studies on data platform decisions, each with its slide deck.
Architecture study, Online marketplace
A PostgreSQL listings ETL took 16 hours to backfill one large brand against a 4-hour SLA, and the brief asked which engine to move to. The answer was a redesign: an entity model in PostgreSQL that hits the SLA for €0 in new infrastructure, with two bigger architectures nested on top for later.
16h → <4h
backfill SLA target
Architecture study, Health technology
Mirror a multi-tenant MySQL source into a queryable analytical store in under 5 minutes — built lean for a startup (~$200/mo) but designed to scale from 150M to 1B+ rows. CDC architecture, data modeling, failure handling, and the trade-offs.
<5 min
source-to-query freshness