Hands-on AI/ML engineering, from code to production: AI agents in Python and TypeScript, fine-tuned models (six on Hugging Face, 149M–27B), and a custom sampler inside llama.cpp. Senior AI/ML engineer with 15+ years in production software, building and running AI systems end to end, from the first prototype to live systems people use every day. Independent consultant since 2021 (VodafoneZiggo, ING, Booking.com, Klarna), now through PrimAxiom (Dutch BV). Built one of Klarna's internal AI platforms, used by 5,000+ employees. Works closely with product and business stakeholders and mentors engineers; earlier engineering lead of a 12-person team at Vamp.io (acquired by CircleCI). Specialties: AI agents and generative AI (GenAI) applications, RAG, LLM fine-tuning, evaluation and LLMOps, and the data and infrastructure underneath. Code and models are public: github.com/bgokden and huggingface.co/Berk.
Open-source work below the API.
Graph memory for AI agents: stores facts, extracts entities and cause→effect relations with small open models, and answers "why" questions with a cited chain of facts. Python library on PyPI, MCP server, and a hosted service (memory.primaxiom.ai). Measured in six languages in a public benchmark (reasongraph-bench), including the experiments that did not help.
Answers many typed questions about a text in one forward pass of a language model, without generating text, and returns calibrated probabilities plus whether the text contains enough evidence to answer. Includes the full training pipeline (data, training, calibration, evaluation); six models from 149M to 27B published on Hugging Face.
Small mDeBERTa models published on Hugging Face with ONNX versions for fast CPU inference: place extraction in 13 languages, and a multilingual cause/effect tagger (causal-span-model) used by ReasonGraph. Related tools: synthspan (synthetic NER training-data generator) and place-extractor-mcp (MCP server that gives AI assistants these models).
Developed custom sampler in llama.cpp enabling strict constrained sampling in LLMs. Library for constrained text generation with Large Language Models for structured output using GGUF models.
Contributor to Spotify's approximate nearest neighbors library for vector search. Built custom implementations for production use cases.
Focus: Machine Learning applied to signal/image/audio processing. Developed classification models using SVMs, including full ML pipeline: data collection, feature engineering, model development, testing, and validation. Technologies: MATLAB, Python, Support Vector Machines. Worked as Research and Teaching Assistant.