Mert Cobanov
Senior AI engineer building agents with complex memory: backend systems, vector databases, diffusion model training, and generative pipelines. Mostly Python, FastAPI, and Pydantic AI.
Writing about AI engineering, homelabs, and whatever I'm tinkering with
Writing
- How I trained an object detector without drawing a single box Building a water bottle detector from scraped images and pseudo-labels, and what the three bugs along the way turned out to teach. TR
- How fast can a 27B model run at home? Running Qwen3.8-27B on four different machines, and what the numbers turned out to actually measure. TR
- KV Cache & Flash Attention LLM inference optimizations, one at a time: a smarter cache, a faster kernel, a better scheduler.
- How AI agent memory works Language models forget the moment they finish replying. Memory is everything the system around them does to make that not matter.
- From latent spaces to JWTs: how agents taught me backend I wrote code every day for years and somehow felt like I wasn't really writing code. Let me explain.
- See all 10 posts
Projects
-
Ahead
A bilingual daily briefing that turns the AI and technology news cycle into a sourced five-minute morning read.
ahead.cobanov.dev
-
Fly Dino
16The original Chromium Dino controlled by a measured fly connectome circuit and a genuinely trained neural readout.
flydino.cobanov.dev
-
Android TV cleanup
63A reversible, no-root guide and ready-made prompt for removing ads and factory clutter from Android TV.
tv.cobanov.dev
-
Instagram Unfollowers
41Find and unfollow people who do not follow you back, directly inside Instagram with no third-party account access.
instagram.cobanov.dev
-
tshare
Drop a file, get a link, gone in 24 hours. A temporary file drop for people, terminals and coding agents.
share.cobanov.dev
-
terminal.army
A multiplayer space strategy game you play from your terminal.
terminal.army
-
herdrchat
3Control your coding agents from your phone, on iOS and Android.
herdrchat.cobanov.dev
-
depth2normal
116Turn a depth map into a normal map in your browser, or estimate the depth of a photograph first. A 1.7 KB ONNX graph, nothing uploaded.
depth2normal.cobanov.dev
Libraries
-
Semantic text clustering using sentence embeddings and agglomerative clustering.
-
Aspect ratio bucketing for diffusion model training. PyTorch native, DDP correct.
-
Unsupervised image clustering with modern deep embeddings, PCA and K-Means.
-
depth2normal 116
pypiConvert depth map images to RGB normal maps for shading and 3D workflows.
-
Summarize web pages and YouTube videos with pluggable LLM backends.
-
Realtime webcam streaming and video processing with a FastAPI UI.
Contact
Based in Los Angeles. Open to interesting projects, collaborations, and long emails about AI, art, and tooling.