Medaş · 2024 · Bronix
Mole
Ground-penetrating radar surveys turned into detected, mapped underground objects.
- PyTorch
- YOLO
- RF-DETR
- YOLO-NAS
- FastAPI
- Next.js
- Supabase
- Label Studio
Open to new roles
Software & AI engineer · previously co-founder at Bronix Software
I design and ship production systems across mobile, backend, and applied AI: factory vision, autonomy, enterprise data products, and consumer software.

User Interface
Web and mobile front ends I ship with:
Backend
Python, Node.js and Go infrastructure, built to scale.
Applied AI
LLMs, vision models, and the apps around them.
30,000+users on yükAl
7deep dives
8studio products
85tools on the bench
Large language models. I love choosing the right model for a job, wiring it into a real product, and making its answers something people can trust.
Vision AI is the other half: teaching models to find things on factory lines, in road scenes and inside radar scans. And I will happily defend LLM wrapper apps: a thin, well-made layer over a strong model is often the fastest way to something people actually use.
Backend infrastructure in Python, Node.js and Go. FastAPI and Express services, Go workers and WebRTC transport, RabbitMQ queues, and data on PostgreSQL, MongoDB, Supabase and Firebase, built to scale rather than to demo.
At Bites Savunma Sanayi I spent eight months as a simulation engineer, building defense simulations in Unity and Unreal with C# and C++.
Work
Deep dives you can open, plus the apps we shipped at the studio.
Radar in
A cart drags ground-penetrating radar along the street. The traces come back as binary files, and I decode them into images a model can read.
Three detectors, one teacher
YOLO, RF-DETR and YOLO-NAS each mark the buried objects. Their boxes are merged into a teacher model that labels the next round.
Section becomes map
The GPS track is cleaned with signal, speed, distance, angle and outlier filters. Each box is projected back onto it as a map polygon with coordinates.
One platform
Upload a survey, get detections, plan around what is underground. A FastAPI detection service and a Next.js map dashboard.
Drawn in your browser from sample data. Not Medaş survey data.
Medaş · 2024 · Bronix
Ground-penetrating radar surveys turned into detected, mapped underground objects.
Freight marketplace · 2024 · Bronix

An LLM turns WhatsApp group messages into structured freight listings that drivers and shippers can actually search.
Bright Machines · Kuartis
A labelling-to-training platform built on open-source CVAT: factory workers label their own line, and the models that count production come out the other end.
First-party · 2026 · Own product
A doctor's practice in one app, with an AI assistant that talks to the records: patients, examinations, prescriptions, lab results and appointments, in English and Turkish.
Rönesans · Bronix
An internal LLM assistant that answers employee HR questions and draws the graphs.
Brisa · 2025 · Bronix
Sales, cycle, and surplus–shortage forecasts computed from the Excel and SAP reports teams already had.
Modo · Kuartis
Embedded iframe payment forms, the payment backend behind them, and security work on a live payment system.
Apps from my Bronix years, each one built end to end.

Walkable
Pick the safer way home. Every street gets a walkability colour and the route sticks to the good ones. No account needed, and your location never leaves the phone.

Toxic Flamingo
A to-do list that refuses to be polite. It nags, roasts and guilt-trips you until the task is done, with a Rage Mode, app blocking and a Pomodoro timer.

NextRoom
Snap a room, a garden or a whole yard, choose one of 12+ styles or describe your own, and the redesign comes back in seconds.

Next2Me
One shared space for two: daily streaks, a drawing board you both sketch on, a couple calendar, 3,000+ questions to talk about, swipe-to-match on films, food and plans, and a private encrypted chat. One upgrade covers both of you.

Banana Labs
An AI studio for UGC videos, product shots and images. Several models run side by side, so you compare takes instead of rerolling one.

LunaGlow
Skincare routines an AI builds around your own skin, with progress photos that show whether they work.
Toolkit
The bench I actually use.
Tools with a dot shipped in a deep dive. Pick one to see where.
Items: 23
Models, inference, and the agent layer around them.
Items: 11
Robots, simulators, and the data loops that train them.
Items: 10
Shipped through both stores, subscriptions included.
Items: 9
Server-first React, typed end to end.
Items: 11
Services, queues, and enterprise sign-in.
Items: 17
Where the systems live and what they remember.
Items: 10
Written daily, not listed for keywords.
AI workflow
Agents do the repetitive half of my week. These are the pieces I actually run.
My projects run from Claude Code with my own skills, so a task starts from a written playbook instead of a blank prompt.
A stop hook holds every session to my definition of done: tests written, run, and read. Design hooks scan UI edits, and permission rules keep each project inside its lane.
I fan work out to parallel subagents, one per review angle or per service, and orchestrate multi-step workflows so a large change lands in one pass instead of twenty prompts.
Cron-scheduled agents keep my apps moving on their own: release and store checks, content jobs, and recurring reports for PatientBook and the studio apps.
Connectors let the agent act instead of advise: Higgsfield for generated image and video assets, Firebase for project and Firestore access, and a browser for live QA.
The same way in Cursor and in Codex: same playbooks, same guardrails, same review loops. The tool changes, the workflow does not.
About
Electrical engineering first, then simulators, robots, and a lot of production traffic.
İhsan Doğramacı Bilkent University
B.Sc. Electrical and Electronics Engineering
TED Ankara Koleji
Simulation Engineer
Defense simulation in Unity and Unreal, written in C# and C++.
Full-stack and AI engineer
Factory vision, autonomy, and the web products that put inference in front of clients.
Co-founder
Co-founded a software studio and built production systems for Rönesans, Brisa, Medaş and others.
Now
I want to build technology that people actually use. Not demos and not slide decks, but apps that tens of thousands of people open every day, like yükAl, with AI quietly doing real work behind them.
Right now I build and run PatientBook: a Flutter app on iOS, Android and the web, with Firebase underneath and an AI assistant beside the records.
I am looking for my next team. If you are building production AI, mobile or backend systems and want someone who can own them end to end, from the model to the queue to the app store, I would love to hear from you.