I build LLM systems that run in production, and the software around them: data model, backend, API, front end, and deployment.
I am an AI engineer in Brussels. I have worked with Python and machine learning for seven years. Most of my work is LLM systems in production. Retrieval quality, evaluation, and failure handling take more time than the model call.
I work across the whole stack: schema, backend, retrieval layer, front end, deployment, and monitoring. I use AI coding tools every day, so a working version reaches the people who asked for it within days.
Before LLMs became common, I worked for years on NLP, document processing, and computer vision. I build these systems through ACA Group today. Earlier I worked at Odoo and, through Exellys, with ING and Contract.Fit.
Most of my work starts from a business process that already exists. I meet the people who do that work, and together we split the process into steps a system can take over.
We go through the process step by step and agree which steps the system handles. Written requirements rarely cover every case, so we settle the rest during the build.
I build the first version in days and give it to the people who will use it, so they can correct it while the system is still easy to change.
Schema, backend, retrieval, front end, deployment, and monitoring. I deploy what I build, and I read the logs after release.
I work in OpenCode and switch between GPT, Claude, and EU-hosted open-source models. The choice depends on the task and on what the client allows to leave their infrastructure. Writing, refactoring, and reviewing code all run through these tools, and they account for most of my speed.
Retrieval quality, chunking, evaluation, data quality, and the handling of failed calls. These take most of the effort on an AI feature, and they decide whether it stays in production.
Deciding what to leave out is part of the work. I propose a scope I can deliver, and I explain what I left out and why.
Each bar uses the same 2019 → 2027 scale. Scroll down to move back in time. Click a project to open it.
A language practice app you use by speaking. It listens, answers, and corrects what you say. I build and run it under CodeBakery.
A full-stack template that other teams use to start AI document projects.
A .NET backend that answers questions from citizens and accepts incident reports. It classifies each reported photo and sets its urgency.
Crawls Belgian municipal websites, structures the text, and answers questions with citations.
Answers questions about client documents and cites the exact passages it used. FastAPI, Celery, and pgvector.
Scores each CV against a job posting with an LLM and explains the score.
Answers support questions from product manuals and cites the exact section.
Matches candidates to job descriptions and explains each score. My own product.
Routes pedestrians over safe crossings. An LLM turns each step into a spoken instruction.
Reads scanned case files with OCR, splits them into separate documents, and searches them by meaning.
Analysts without technical training run LLM text analysis over country-risk documents and read the results.
YOLOv9 object detection on Jetson Orin units in garbage trucks, managed as a fleet with AWS Greengrass.
A microservice stack with authentication, caching, messaging, and a database already connected. New projects start from a working baseline.
Python data products that run in a Kubernetes data mesh shared by several government departments.
A Django tool that plans field appointments. It runs on the client's existing Navision and Active Directory data.
Fine-tuned GPT-Neo on horoscopes collected from the web and served it as an API. My first generative model in production.
Rebuilt a product scraper. It crawls search results and extracts products with XPath, page metadata, and DOM-tree similarity.
My consultancy company. Client AI work runs through it, and I build Speakroom and Swiftalent under it.
Features and performance work on the open-source website builder, in Python and the JavaScript/OWL front end.
Moved machine-learning models to Google Cloud with their data processing, deployment, and scheduling.
Classifies invoices and receipts and extracts their fields. Includes Dutch language processing.
An EDSR implementation that upscales any image.
Document classification on image and text features together. State of the art on RVL-CDIP and Tobacco3482.
I take work on AI products and the systems behind them: retrieval, backends, and internal tools.
EMAILcontact@kamen.be LINKEDINin/kamen-zhekov ↗