Projects CV
Back to Projects View on GitHub

Machine Learning Operations

MLOps Pipeline

A lightweight real-time machine learning pipeline that collects user input, stores data, retrains models, and serves predictions as new data arrives. It combines scikit-learn kernel density estimation with a PyTorch-based neural network density regressor, with FastAPI orchestrating the end-to-end workflow. Try the demo below!

Workflow

User inputData point
DjangoWeb application
FastAPIML API
SQL databasePersisted data
FastAPI engineOrchestrate Training
ML enginescikit-learn + PyTorch
DjangoClient updates
PlotlyLive visualization

Docker · Uvicorn · VPS

Input points

Drop points on the canvas and compare a kernel estimator with a neural density model.

Checking the density service…

Clear canvas

Kernel density

Neural density

Berk Can Özmen

+49 xxx xxx xxxx | mail@bcozmen.com | bcozmen | bcozmen.com | bcozmen
Download PDF

Summary

M.Sc. Machine Learning Researcher with a strong background in deep learning, representation learning, reinforcement learning, and computer vision. Experienced in developing and evaluating learning algorithms, designing experimental frameworks, and integrating visual perception & decision-making into autonomous systems.

Skills

Programming: Python, C++
ML Paradigms: Deep Learning, Reinforcement Learning, Representation Learning, Meta-Learning
ML Architectures: Transformers, CNNs, Graph Neural Networks, Diffusion Models
Computer Vision: Geometric Deep Learning, Neural Radiance Fields (NeRF), 3D Reconstruction, OpenCV
ML Frameworks: PyTorch, Scikit-learn, TensorBoard, HuggingFace, Gymnasium
Infrastructure: Linux, Git, Docker, Kubernetes, AWS
Sci. Computing: NumPy, SciPy, Pandas, Numba, CUDA
Languages: Turkish (Native), English (C2), German (C1), French (A2)

Experience

06/2025 - Present
Berlin
Independent ML Engineer bcozmen.com
Self-Employed
  • Architected and built an end-to-end AI-powered knowledge management system, including semantic search over 100M+ records, RAG pipeline, real-time collaboration, and an integrated AI assistant.
  • Built end-to-end ML applications and numerical simulation pipelines, covering model implementation, data processing, API development, and cloud deployment.
09/2023 - 09/2024
Berlin
Machine Learning Engineer
Fraunhofer SIRIOS
  • Developed a reinforcement learning-based method (SAC & DQN) for ambulance and fire truck allocation, to optimize for mortality risk, response-time weighted for urgency, and operational costs.
  • Implemented the learning algorithm and reward function, validating learned policies against empirical probability distributions in tractable scenarios and comparing their performance with greedy allocation baselines.
  • Evaluated policies across hand-crafted scenarios for complex cases, assessing compliance with operational regulations and performance on task-specific metrics.
09/2021 - 04/2023
Berlin
Computer Vision & Robotics bcozmen.com/scioi
Science of Intelligence
  • Developed a computer vision pipeline to reconstruct 3D geometry of a lockbox puzzle from camera observations mounted on a robotic arm.
  • Integrated the reconstructed geometry into an internal world model, which guided the robotic arm to collect additional viewpoints and iteratively update the model.
  • Reconstructed the lockbox geometry and determined a feasible robot approach configuration for subsequent kinematic exploration.
09/2019 - 06/2020
Berlin
Internship at Neural Information Processing Group
Technische Universität Berlin
  • Implemented a simulation environment for FitzHugh-Nagumo based resting-state brain network models, using empirical structural connectivity and functional connectivity data.
  • Explored the parameter space of the FitzHugh-Nagumo model using genetic algorithms and Bayesian optimization with Gaussian processes on a distributed computing setup (HPC).
09/2017 - 04/2019
Berlin
Software Developer
Ten8
  • Developed and maintained the company's website and internal application for a Berlin-based startup.
  • Developed a web scraping tool to collect and process initial coin offering (ICO) data, automatically alerting the team to new and potentially interesting opportunities.
2013 - 2016
Istanbul
FRC Team Member & Leader
Inanc Mechatronics Club
  • Designed and built competition robots for the FIRST Robotics Competition (FRC), leading a 10-member technical team in 2015.
  • Built robots end-to-end, spanning 3D CAD, fabrication, machining, mechanical assembly, and programming for autonomous and teleoperated operation.

Projects

2024 - Present
Deeppy bcozmen/deeppy
  • Created a modular PyTorch training framework for rapid prototyping and experimentation in deep learning research.
  • Decoupled model architectures, data pipelines, optimizers, schedulers, and training logic, with integrated logging, visualization, checkpointing, and automatic mixed-precision training.
  • Automated device placement and tensor management across GPU/CPU memory spaces, with data buffering to support large datasets.
  • Continuously improved through practical use across 5+ projects and my master's thesis.
  • Includes DQN, SAC, Beta-VAE and transformer-based architectures as tutorials.
2022 - Present
Agent-Based Neuro-Evolution bcozmen/abne
  • Developing a physics-based multi-agent environment for studying neuroevolution, where agents evolve neural network controllers through selection rather than an explicitly defined loss function.
  • Investigating compact representations of agent behavior using message-passing graph neural networks, alongside non-destructive crossover and mutation operators for evolving neural controllers.
  • Exploring how evolutionary selection pressures drive the emergence of complex, adaptive behaviors in multi-agent systems without explicitly designing the behaviors to be learned.
2025
r-hash-SANE bcozmen/r-hash-SANE
  • Extended the SANE architecture to hash-grid NeRFs, developing a transformer-based autoencoder for learning latent representations of NeRF parameters.
  • Implemented and evaluated reconstruction and contrastive learning objectives, successfully learning latent representations; explored rotation-related learning objectives, limited by the available training data.

Education

2026
Berlin
M.Sc. Computer Science, est. 1.3 (3.8/4.0) bcozmen/master_thesis
Technische Universität Berlin
  • Thesis: Interpreting NeRF Weights Across Architectures with Graph Meta-Networks for Geometry and Appearance Attribution
  • Supervisor: Prof. Dr. Olaf Hellwich
  • Proposed a self-interpretable Graph Meta Network with learned gating and attention-based pooling to identify task-relevant substructures in NeRF weight spaces.
  • Evaluated the approach on geometry and appearance prediction using NeRF representations paired with corresponding RGB and geometry histograms, demonstrating that the model identifies relevant computational subgraphs.
2021
Berlin
B.Sc. Computer Science, 1.7 (3.5/4.0) bcozmen/bsc_thesis
Technische Universität Berlin
  • Thesis: Modeling Functional Connectivity in Health and Psychiatric Disorders | Thesis Grade: 1.3
  • Supervisor: Prof. Dr. Klaus Obermayer