I build machine learning and LLM systems that keep working after launch: pipelines, training, deployment and monitoring on Azure and GCP. Syntonic is the independent practice of Ilhan Karić, a senior machine learning engineer working remotely across Europe.
Typical engagements: a proof of concept, a model taken into production, ongoing work alongside a team, or an existing system made reliable.
Models built from the data up: pipelines, training and optimization, then monitoring for regression and drift for as long as they run.
Assistants and agent workflows on frontier models, taken from proof of concept to production, with evaluation to show they work.
Deployment and operations on Azure and GCP: containers, automated builds, cost alerts, and the monitoring to know when something breaks.
The product around the model when a project needs one: backend services, APIs and frontends, held to the same standard as the model.
Running a project from roadmap to delivery, mentoring engineers as a team grows, and research when a problem calls for it.
I am Ilhan Karić, a senior machine learning engineer and the owner of Syntonic, an independent practice registered in Zenica, Bosnia and Herzegovina. I live and work in Mostar. I have been writing software since 2007, first as a hobby, then through school and competitive programming, and since 2018 as a profession.
I have eight years of professional experience across production systems, applied research and university teaching. My work covers the whole life of a machine learning product: I acquire the data through scraping and ETL, do the data science needed to understand it, then build, train and optimize models and deploy them with monitoring for regression and drift. In recent years the focus has shifted to LLM and agentic systems built with LangChain and the Pydantic ecosystem, running on Azure and GCP with models from Anthropic, OpenAI, Google and xAI.
Since September 2026 I work at ZenDev, bringing AI into the daily work of a large Swedish retailer. Before that I spent three and a half years as a senior machine learning engineer at evoila Bosnia, part of the German evoila group, working for clients across Europe, including a year leading a custom AI assistant from proof of concept to production. Earlier still I taught Software Engineering, Software Architecture and Programming at the Faculty of Information Technologies in Mostar and did research in artificial intelligence.
Dates, education and publications are all in my full CV, available as a PDF.
An AI assistant for a European software company, which I led for more than a year from proof of concept to production. Roadmap and project management alongside daily work on the backend, API layer and frontend.
Models that had to keep running long after delivery. Data acquisition through scraping and ETL, training and optimization in TensorFlow, deployment on Vertex AI and the Azure AI services, and the monitoring for regression and drift, with cost alerts, that kept them running.
Assistants and agent workflows on LangChain and Pydantic AI, evaluated to show they work before they ship. Backends in Python and C#/.NET, builds automated through Docker and GitHub Actions, and the frontier model chosen for the task rather than the vendor.
A 3D mass monitoring system tracking upwards of 250,000 people in real time, for high-security facilities, event venues and oil rigs. Built in C# on the Unity3D engine at Four Game Studio in Mostar.
A research prototype of an AI teaching assistant for university courses, published at ICCAIRO 2017. The most cited of my papers, with sixteen citations to date.
Four years at the Faculty of Information Technologies of the University “Džemal Bijedić” of Mostar, teaching, grading and supervising students, alongside research on neuroevolution and similarity search.
Designing and building AI solutions on Azure, from the Azure AI services to generative AI applications.
Verify on Microsoft LearnA short note about what you are building is enough to start. I work remotely with teams across Europe and beyond, in English, German or Bosnian, and keep Central European working hours.
Typical engagements are a proof of concept for a machine learning or LLM feature, taking an existing model into production, or joining a team for ongoing work. Scope and duration are flexible, from a few weeks to a standing arrangement.
Existing systems are welcome too: a pipeline that needs monitoring, a model that has drifted since launch, or an assistant that has to become reliable before it ships.
The first conversation is about the problem, not the tooling. Once the goal, the data and the constraints are clear, you get a written proposal with scope, milestones and a price before any work starts.
Everything is documented as it is built, so nothing depends on one person, and every engagement ends with a proper handover.