AI Engineer
Terya (TeamSystem Group)
Terya makes management software for retail, food service and the food industry. I work on the AI side of its LLM-based systems: retrieval, prompts, tool calling, evaluation and deployment. These are the main ones.
Chatbot on hundreds of technical manuals
A chatbot for the service desk that answers support questions from several hundred technical manuals and documents, so an operator gets the right procedure without searching for it.
The manuals are split into clearly delimited paragraphs following an internal procedure, which makes retrieval more accurate. Every question is searched twice, by meaning and by keywords, and the passages found go through a final ranking. The answer is detailed and precise, and it is built only from what the manuals say, to avoid hallucinations.
Speech-to-speech voice agent for support calls
A speech-to-speech agent that answers support calls in real time. The caller speaks naturally and can interrupt it at any moment, while the agent collects the information it needs and saves it in the ticketing system.
Code knowledge agent
A knowledge base built on the company's codebase: the code is indexed and turned into documentation, so that questions about it can be asked in plain language.
The documentation and the processed code then serve as context, as in a RAG system, to make development work faster: more context for each task, and fewer tokens spent on paid models. Answers also adapt to who is asking and to the type of request, such as an impact analysis or documentation.
Tool-calling orchestration for the ERP
An orchestration layer on top of the ERP: people ask in plain language, and the model picks and calls the tools that read or change the data, without anyone needing to know the tables behind it.
The tools are not hard-coded but discovered at request time, and they cover the everyday operations on tables and business objects: viewing, adding, editing, counting and deleting records. Every deletion asks for a confirmation first.
Self-hosted inference
An NVIDIA DGX Spark on the premises runs all the models, language, embedding and reranking, instead of a cloud API, so internal data stays inside the company network. The language models are from the Qwen family and are served with vLLM and Ollama. All the systems here use it, with cloud models as a fallback.
Second-level support agent in development
An AI agent for second-level support that works out dynamically how to use the information it is given. For each request it picks the sources and tools it needs, among tickets, knowledge bases, logs, the ERP and the code knowledge agent, and combines what it finds into one answer with its sources.