Backend engineer. Python, mostly. I care about the parts of a system that fail quietly: message delivery, caches, retries, the second request that does not look like the first.
Open source
Merged upstream:
- CPython · Fix
asyncio.as_completed()outside a running event loop - CPython · Remove dead
_repr_info()methods fromasyncio.windows_events - Django · Include the invalid view type in the
path()/re_path()error - Django · Reorganize the security policy page
- Django · Fix outdated docs for the
JsonResponsesafeparameter
What I build
- txoutbox · transactional outbox relay for asyncio.
Bring your own table and your own broker; the core has no dependencies and ships contract tests
for adapters.
pip install txoutbox - promptcachelint · explains why an LLM prompt cache missed.
Diffs consecutive requests and names the block, the offset and the bytes that broke the prefix,
for Anthropic and OpenAI. Zero dependencies.
pip install promptcachelint
How I work
- Contracts first, then code. Small typed interfaces, adapters behind Protocols, a contract test suite that any adapter has to pass.
- Precision over promises. If a tool reports a number, the README says whether it is measured or estimated.
- Reviews are input. Every finding gets reproduced, then fixed or explicitly declined with a reason.
- Human plus agent. Much of this code is written with an AI coding agent, and the commits say so. The agent works under fixed rules: a failing test before a bug fix; lint, types and tests before every commit; an independent review pass on non-trivial diffs; nothing is pushed without my explicit approval.
Elsewhere
Notes: osmnvaslan.github.io
PyPI: txoutbox · promptcachelint
Also: batch-creative, a small Next.js app for batch image generation
Mail: osmnvaslan@gmail.com



