Notebook of an AI engineer · May 16, 2026
Hande
Kafkas.
I build and write about LLMs and agents.
AI engineer and technical writer. I build the systems behind language model products — and write about what survives contact with production.
Currently
AShipping
A planner/executor runtime for agentic systems where the plan is a first-class artifact you can read, replay, and diff.
BWriting
A short series on the patterns that survive between agent demos and production. One post out, two in drafts.
COpen to
Consulting engagements with teams building serious LLM products. Especially the parts that aren't the model.
- May 10, 20263 min
How agents actually plan, in practice
Notes from six months of shipping production agents — and why the textbook ReAct loop is almost never what you want.
- Apr 22, 20262 min
Notes on long context windows
A million tokens does not mean a million useful tokens. A short field guide to what context windows actually buy you, and where to spend the budget.
2026 · Lead engineer & technical author
Supervisor
A two-tier planner/executor architecture for agentic systems, with first-class plans, replay, and step-level observability.
2025 · Solo build
Groundtruth
A reproducible eval harness for RAG systems: golden questions, citation-level scoring, and regression detection between model and index changes.
2025 · Co-founder & maker
Field Notes
A markdown-first writing app that turns scattered engineering notes into structured drafts, with model-assisted outline extraction and tone preservation.
Diff Summarizer
Tiny prompt that turns a `git diff` into a one-paragraph plain-English summary aimed at non-engineers. Surprisingly hard to tune.
wipTone Mirror
An experiment in style-transfer for writing: paste any text, get back a version in the voice of three different reference authors. The failure modes are the interesting part.
Building with LLMs in earnest? I’d like to hear about it.