Hi, my name is
Yusuf Serhat Özkan.
Embedded ML & backend engineering.
I'm a Computer Science student at Maastricht University. I work on machine learning for ARM microcontrollers, robotics, and cloud backends.
About Me
I'm a Computer Science student at Maastricht University, graduating in 2026. My main interests are embedded machine learning, robotics, and backend engineering.
For my bachelor's thesis, I deployed a spiking speech-enhancement network on an STM32 microcontroller. My other projects include robot motion planning, encrypted machine learning, and a tutoring platform backend.
I work mainly with Python, C, C++, and Java.
Projects
-
Bachelor Thesis
Spiking Neural Networks on Embedded Hardware Code
Feb–Jun 2026
I got a spiking speech-enhancement network running end-to-end on an STM32U585 microcontroller (ARM Cortex-M33). The stock model overflowed the chip’s SRAM by 1.82×. The weights were fine; the real cost was the decoder’s overlap-add reconstruction, so I rewrote it as a single-frame streaming graph that produces identical output.
Peak memory cut 30.5× (1.37 MiB → 46 KB) · 17.42 dB SI-SNR · near real-time on-device
-
Robotics Project
Dynamic Collaborative Motion Planning
Oct 2025–Jan 2026
A motion-planning pipeline for a KUKA iiwa 14 cobot that shares its workspace with people. I benchmarked sampling-based planners (RRT*, PRM*) with and without shortcutting, then added closed-loop, time-sliced execution because the controller gives you no way to interrupt a trajectory once it starts.
9,000 planning trials · 100% success · ~10× fewer trajectory waypoints via shortcutting
-
Decentralized Medical Data Sharing
Oct–Dec 2025
A Solidity platform that lets patients decide who can read their medical records and for how long, with the option to revoke access at any time. Every access attempt lands in an on-chain audit trail; the records themselves stay encrypted off-chain.
6 core contracts · 50 passing tests · gas-benchmarked with Hardhat & Foundry
-
Elevate Tutoring Platform
Sep–Dec 2025
The backend for a tutoring platform: Spring Boot with OIDC/OAuth2 login, MinIO for S3-compatible object storage, everything running in Docker. Terraform provisions the Azure infrastructure and a 4-stage GitLab pipeline handles CI/CD.
OAuth2 · Azure + Terraform · 4-stage CI/CD
-
Encrypted ML Inference for Secure Healthcare Code
Feb–Jun 2025
Machine learning on medical data that never gets decrypted. I ran a neural network entirely under homomorphic encryption (Microsoft SEAL, Concrete-ML) and measured what that costs: how much accuracy you keep and how much latency you pay compared to plaintext inference.
85.5% accuracy parity, plaintext vs fully-encrypted inference
-
PPO vs. SAC in Unity ML-Agents
Sep 2024–Jan 2025
A head-to-head benchmark of PPO and SAC in Unity ML-Agents, on one continuous task (Crawler) and one discrete task (Push Block). I compared convergence speed, final reward, and CPU cost to figure out when each algorithm is worth deploying.
PPO: faster convergence & lower CPU · SAC: stronger exploration, ~2× training time
Skills
Embedded & Hardware
AI / ML
Programming
Backend
Cloud & DevOps
Robotics
Security
Spoken Languages
Education
Maastricht University
2023–2026
BSc Computer Science
Coursework
05.
Contact
I'm open to internships and new-grad roles. You can reach me by email or on LinkedIn.