I build local-first AI tools, ML systems, edge AI projects, and public desktop products with a focus on reliability, observability, and real-world system behavior.
A recurring theme in my work is:
Deterministic Decision Making in Non-Deterministic Environments
Real systems are noisy. Inputs drift. Timing breaks assumptions. Hidden state appears.
I am interested in building systems that are not only accurate in a notebook, but also inspectable, reproducible, debuggable, and usable after something goes wrong.
My projects often sit between:
- machine learning and software engineering
- local AI tools and real user workflows
- edge devices and constrained inference
- data collection and model behavior
- research prototypes and usable systems
| AI Workflow Reliability | Prompt versioning, local evaluation, debugging utilities, and inspectable AI workflows. |
| Local AI & Developer Tools | Small practical tools for local experimentation, document workflows, and developer productivity. |
| Edge AI & Embedded ML | TinyML, Raspberry Pi Pico projects, sensor data, firmware, and constrained inference. |
| Computer Vision & Data Tooling | Dataset collection, sanity checking, real-time inference, and visual debugging workflows. |
| Project | Description | Links |
|---|---|---|
| OpenAnima | Open-source desktop overlay engine for Windows. Built as a local-first desktop visual companion and creative overlay system. | GitHub · Website · itch.io |
| PromptLedger | Local-first prompt version control with prompt history, labels, diffs, metadata, and a lightweight workspace dashboard. | GitHub · PyPI · Series |
| DataLens | Read-only image dataset sanity checker for detecting corrupted images, duplicates, label issues, and silent data problems before training. | GitHub · Write-up |
| JumpNet | Vision-based game agent focused on data collection, timing, real-time inference, and deployment behavior. | GitHub · Write-up |
| Edge AI Sound Classifier | TinyML audio classifier running on Raspberry Pi Pico-class hardware. | GitHub · Series |
I write technical essays, project series, and failure-analysis notes about the systems I build.
| Essay / Series | Theme |
|---|---|
| Deterministic Decision Making in Non-Deterministic Environments | The engineering philosophy behind my projects |
| I Intentionally Built a Bad Decision System | Hidden state, side effects, and decision-system failure modes |
| PromptLedger Series | Local-first prompt versioning and prompt workflow tooling |
| OpenAnima v1 | Public desktop product release note |
| DataLens | Dataset inspection and ML data reliability |
Full writing archive: dev.to/ertugrulmutlu
From Human-in-the-Loop to Agent-in-the-Loop
A public DataTalks.Club talk about how ML workflows evolve into agentic feedback systems, and where human judgment, monitoring, evaluation, and system design still matter.
I am a Computer Engineering student at RWTH Aachen University and a Werkstudent Researcher at Fraunhofer IAIS.
I like building projects that sit between research ideas and practical systems: prototypes, tools, dashboards, edge deployments, and workflows that can be tested outside clean notebook environments.
I prefer systems that can be:
- inspected
- reproduced
- logged
- debugged
- improved after failure
Because a demo working once is not the same thing as a system being reliable.