Muhammed Ali Yıldırım
Menu

Hi, I’mMuhammed Ali Yıldırım

From data to decisions.

I turn AI models into systems that work in the real world.

Applied AI / ML Engineering · Computer Vision · Edge AI · Research / R&D

Explore the workCVContact
01Bachelor’s ThesisComputer Vision · Edge AI · Drone

01Edge AI · Thesis · Two-person team

AI-powered search & rescue system

The model works.And the system?

An IMX477 camera on a drone, YOLOv11 running on a Jetson Nano, and a DeepStream RTSP video stream: an end-to-end system, tested in the field, that marks people and leaves the final decision to the operator.

My contribution

The system was built end to end by a team of two—from research and procurement to sponsorship and assembly. I was primarily responsible for the computer vision, dataset and model development, training and evaluation, and the Jetson-based edge AI pipeline. The project was supported by Gazi University’s Scientific Research Projects unit (BAP) and has been completed.

  1. IMX477 camera
  2. Jetson Nano · YOLOv11
  3. DeepStream · RTSP
  4. Ground station

Selected model: YOLOv11

Precision 0.9109 · Recall 0.8065 · mAP@0.5 0.8694 — PROJE dataset, end-of-training validation metrics

Show the two pathsBack to the footage

Perception and flight, two separate paths.

Perception: the camera image is processed on the Jetson Nano; detections are drawn onto the video and streamed over RTSP to the ground station.

Flight: flown from an RC transmitter through a Pixhawk 6C, and semi-autonomous through safety functions such as return-to-home (RTH) and automatic take-off. No command is sent from the Jetson to the flight controller.

Read the case studySource code
02TÜBİTAK 2209-AResearch
DualViT–CNN + ArcFace 2209-A · EXPLANATORY ARCHITECTURE SIMULATION INPUT 01 · Data CNNLocal features ViTContext features 02 · Model Cross-attention Training with ArcFace 03 · Comparison

02Research · TÜBİTAK 2209-A

Identifying PCB materials from thermal images

To understand a result,understand the experiment.

A model learning from simulation data. A hybrid architecture connecting local detail with the wider image.

My contribution

I was the project lead and the machine learning and computer vision work was mine: I designed and trained the DualViT–CNN + ArcFace hybrid architecture and compared it with classical models.

  1. Simulation data
  2. CNN + ViT
  3. Comparison
  4. Evaluation

Published · Gazi University Journal of Science Part A, 2026

AI-Assisted Thermal Response Classification of PCB Materials for Energy-Efficient Femtosecond Laser Processing: A Simulation-Based Study

Unfold the experimentSee the experiment as a whole

Three layers, three responsibilities.

Where the data begins — Simulations produced the thermal responses of materials under laser processing. I handled image preprocessing and dataset preparation.

How the model is built — I designed and trained the DualViT–CNN + ArcFace hybrid architecture and ran the model experiments.

What the result supports — I compared the model with other learning approaches. Physical system validation is outside the scope of this study.

ScopeCollaborative simulation study. Findings are based on simulated data; no claim of physical validation.

Read the case studyAcademic record
03LabelMateSoftware

03Independent software · Python / PySide6

LabelMate

Data stays local.People make the call.

The model proposes boxes; a person reviews and corrects them, and the result becomes a YOLO-format label. Local inference, editable boxes, undo history and a keyboard-driven workflow: an MIT-licensed desktop application I built on my own.

My contribution

Designed and built on my own, to keep labeling local and offline—protecting the data—and to make the process faster. Its first version grew out of my workplace training at Boğaziçi Savunma.

  1. Model proposal
  2. Human decision
  3. Undo history
  4. YOLO label

Open source · MIT licence

Python · PySide6 · Ultralytics YOLO · Windows setup script

Show the decisionBack to the recording

The model proposes; a person decides.

A proposal arrives with a box, class and confidence; a person deletes, corrects or keeps it. Every saved box is one line, normalised to the image.

YOLO line · class / centre / size0   0.512300 0.438200   0.034100 0.028900
Read the case studySource code
04Boğaziçi SavunmaWorkplace training

04Boğaziçi Savunma Teknolojileri · Workplace training

Real-time detection & tracking system

One output.A much wider system.

Two projects in one internship: a two-stage detection system that tells drones from birds, and detection, multi-target tracking and single-camera distance estimation for fixed-wing UAVs.

My contribution

Field data collection and preparation, model training, two-stage inference, TensorRT FP16 optimization, ByteTrack tracking, distance estimation and the interface.

  1. Field data
  2. Detection + classification
  3. TensorRT FP16
  4. Tracking + distance
See both projectsBack to the recording

Two projects in one internship.

Fixed-wing UAV — YOLO26L with fine-tuning, ByteTrack identities, a detection console and approximate distance from one camera.

Drone vs Bird — detection, cropping and three-class classification; TensorRT FP16 from 12.8 to 26.0 FPS.

ScopeThe training data is confidential under contract; the images on these pages come from open-source data. Measurements were taken on a desktop GPU.

Explore the fixed-wing UAV caseThe Drone vs Bird caseSource code
05TÜBİTAK RUTEInternship

05TÜBİTAK RUTE · Internship

TÜBİTAK RUTE internship

Detail in an image.Structure in three dimensions.

Brake pad surface defect segmentation and railway environment analysis with Rail3D point clouds. Two projects I developed with different data types during the same internship.

My contribution

I carried out the annotation, model experiments, processing workflows and result visualization from start to finish.

  1. Annotation
  2. Model experiments
  3. Processing flow
  4. Visualization
See both projectsBack to the output

Two data types, two analysis workflows.

Brake pads — Pad masks, ROI extraction, stain and crack segmentation, and pixel-based thickness output.

Rail3D / Open3D — HMLS point-cloud processing, feature extraction and machine-learning classification.

ScopeInternship prototypes. Pad thickness is measured in pixels; the point-cloud work uses the HMLS subset of Rail3D.

Explore the brake pad caseExplore the Rail3D case

University · 06 — ∞

And countlessuniversity projects.

From electronic circuits to PCB design and microcontroller programming: throughout my undergraduate studies at Gazi University I built countless small-scale projects.

  • Electronic circuits
  • PCB design
  • Microcontroller programming

Experience & education

Experience

Different settings. The same curiosity.

Boğaziçi Savunma TeknolojileriComputer Vision / Edge AI · Workplace trainingFebruary — June 2026
Contribution and scope

I worked end to end, from collecting field data to a real-time application. I trained two-stage detection and classification models, optimized inference with TensorRT FP16—from 12.8 to 26.0 FPS on a desktop GPU—and built multi-target tracking with ByteTrack, single-camera distance estimation and the application interfaces.

Model development / Optimization / Software

See the work above
TÜBİTAK RUTEImage Processing & Deep Learning · InternshipJuly — September 2025
Contribution and scope

I worked with two kinds of data on railway components. For brake pad images I developed an analysis pipeline of YOLO-based detection, region-of-interest extraction, stain and crack segmentation and pixel-based thickness measurement. For Rail3D point clouds I used Open3D for processing and feature extraction, with LightGBM for point-level classification.

Image processing / Deep learning / Point clouds

Explore the internship work
Education

Gazi University

BSc, Electrical & Electronics Engineering · 2026 · GPA 3.58 / 4.00

CV & professional summary

Contact

Let’s buildthe next system together.

As an engineer focused on turning AI models into systems that work in the real world, I’m open to new opportunities and technical collaborations. If you work on computer vision, edge AI or applied machine learning, I’d be glad to connect.

LinkedIn GitHub Curriculum vitae

Muhammed Ali Yıldırım

Applied AI / ML Engineering

Back to top