Gökdeniz Gülmez - ML Researcher

About Me

I'm an ML researcher and engineer passionate about building practical, production-grade AI systems. My focus is on language model fine-tuning, reinforcement learning, and efficient inference on Apple Silicon. I contribute to the MLX ecosystem and maintain mlx-lm-lora, used by companies like Apple, IBM, Bosch, and Mercedes-Benz.

At Computacenter, I develop time series forecasting systems and analytics infrastructure for enterprise-scale machine learning. In the open-source community, I've published research on neural weight modification (Gabliteration) and released the JOSIE model family. Currently exploring mechanistic interpretability and controllable model behavior through systematic ablation.

Based in Stuttgart, Germany. Enthusiast of anime and violin.

Research Interests

Natural Language Processing
LLM fine-tuning, instruction tuning, alignment techniques (RLHF, DPO, GRPO), and efficient adaptation methods like LoRA/QLoRA on both cloud and edge devices.

Reinforcement Learning
Policy gradient methods, reward modeling, and deep RL approaches for language model alignment and optimization.

Time Series & Forecasting
Advanced architectures for multi-horizon forecasting, temporal fusion transformers, and sequence modeling for complex patterns.

Machine Learning Infrastructure
Apple Silicon optimization with MLX, distributed training, model quantization, and efficient inference for production systems.

Representation Learning
Embeddings, dimensionality reduction (UMAP), and learned representations for semantic understanding and analysis.

Open-Source Research
Practical implementations of cutting-edge techniques, community-driven development, and reproducible research tooling.

Open-Source & Research

mlx-lm-lora

Apple Silicon LLM Fine-Tuning

Production-ready package for LoRA/QLoRA/full fine-tuning on Apple Silicon using MLX. Supports GRPO, DPO, ORPO, CPO training modes with synthetic dataset generation pipelines. Used by Apple, IBM, Bosch, Daimler Truck, Mercedes-Benz.

MLX LoRA RLHF OSS

Gabliteration

Neural Weight Modification Framework

Principled multi-directional neural weight modification technique with theoretical guarantees for selective behavioral alteration in LLMs. Published on arXiv (2412.06527) with automated CLI tool.

Research Published Transformers

JOSIE Model Family

Fine-Tuned LLMs (Qwen3, Qwen2.5, Vision)

Extensive family of instruction-tuned and abliterated models across multiple architectures. Available in multiple sizes (0.5B-31B) with quantized variants. Paper published on arXiv (2512.18901).

LLMs Fine-Tuning Hugging Face

Local-NotebookLM

PDF-to-Audio Podcast System

End-to-end system converting PDFs into natural podcast-style audio. Includes FastAPI backend, Gradio web UI, and Docker deployment. Published on PyPI and GitHub.

Audio FastAPI Production

MLX-LM-LENS

LLM Interpretability & Mechanistic Research

Abliteration and mechanistic interpretability tools for researching and modifying internal representations of LLMs. Enables systematic study of model behavior and steering on Apple Silicon.

Interpretability MLX Research

MLX Ecosystem Contributions

Official Core Contributor

Core contributor to mlx, mlx-lm, mlx-vlm, and mlx-examples. Added support for 20+ LLM architectures. Officially recognized in mlx-lm acknowledgments. Featured in Apple WWDC 2025.

MLX Core Contributor Open Source

Work @ Computacenter

Proprietary systems and internal ML infrastructure (details confidential)

Time Series Forecasting

Production ML System

Advanced forecasting systems for enterprise metrics. Implements state-of-the-art architectures with focus on production reliability and real-time inference at scale.

PyTorch Enterprise Production

Analytics & Embeddings

Semantic Understanding

Large-scale embedding pipelines for semantic analysis and categorization. Dimensionality reduction and visualization for business intelligence dashboards and system monitoring.

Embeddings UMAP Power BI

Cloud ML Infrastructure

Scalable Backend Services

End-to-end ML platform for production services. GPU management, dataset orchestration, and model deployment across cloud environments. Azure Functions and workflow automation.

Azure Functions Cloud GPU

Journey

2025 – Present

ML Researcher & Engineer @ Computacenter

Building production ML systems with focus on RLHF, time series forecasting, and analytics infrastructure. Active in MLX ecosystem; paper on neural weight modification (Gabliteration) published on arXiv.

2024 – 2025

Open-Source & Research

Released JOSIE model family with paper on arXiv. Extended mlx-lm-lora with batch generation, concurrent inference, GRPO, ORPO, CPO support. Created Gabliteration framework for controlled model modification.

2023 – 2024

MLX Ecosystem Core Contributor

Added 20+ LLM architecture support (Mamba, MiniCPM, GLM, DeepSeek, Helium). Implemented DPO, ORPO, GRPO training in mlx-lm. Featured in Apple WWDC 2025. Officially recognized by MLX team.

2023

mlx-lm-lora Package Launch

Created open-source package for efficient LoRA/QLoRA fine-tuning on Apple Silicon. Now widely adopted by Apple, IBM, Bosch, Daimler Truck, Mercedes-Benz. Published on PyPI and GitHub.

2023

Apprenticeship @ Computacenter

Fachinformatiker für Systemintegration (IT Systems Integration) at Computacenter AG, Stuttgart. Transitioned to full-time ML Engineer/Researcher role. Building enterprise ML infrastructure and forecasting systems.

2021 – 2023

IT.Schule Stuttgart

Technical education in information technology and systems integration. Foundational training in IT infrastructure, networking, and systems management. Location: Stuttgart, Germany.

Technical Skills

Core ML Frameworks

PyTorch

MLX (Apple Silicon)

Transformers

Hugging Face Ecosystem

Specializations

LLM Fine-Tuning (LoRA/QLoRA)

RLHF (DPO, GRPO, ORPO, CPO)

Time Series Forecasting

Model Interpretability

Enterprise & Deployment

Azure Functions

Power BI

Cloud GPU Infrastructure

Model Quantization

Tools & Infrastructure

Python & Data Science

Git & Open-Source Dev

FastAPI / Gradio

Research & Publishing