Open role in Cluj-Napoca
AI Engineer
Build the retrieval, agents and tool connections that put models to work on the customer's own data. Measure every change against questions from the people who use it.
The role
Make the model useful on real data, and prove it.
Pexon Romania is hiring an AI engineer in Cluj-Napoca to put LLMs to work on customer data: retrieval over SAP, MES and PLM systems that keeps their permissions, MCP servers and agents that call those systems safely, and evaluation sets that prove a change is better. Jobs are based at the Cluj office.
A person behind the process
Meet Laura Deliu.
HR Business Partner, recruiting

Laura runs recruiting for the Cluj hub. Your application reaches her first, she answers your questions about the process and the range, and she brings in the engineers you would actually work with. Later in the process you also meet Noel Dinger, who runs the company.
The work
What you will do.
- Tune chunking, embeddings and ranking, and carry the source system's access rights into every query
- Build MCP servers and connectors with narrowly scoped tools and the user's identity carried through each call
- Build evaluation sets from real questions and failures, and run them on every change, on hosted and private models alike
What you need
Experience and skills.
Production software
Backend or full-stack engineering experience, with a system in production you can discuss in detail.
LLM work you shipped
Retrieval, tool calling or agents that real users relied on, and what you changed when they got it wrong.
Python in depth
Most retrieval, agent and evaluation code is Python. Customer systems decide the rest.
Cloud and containers
Working experience with AWS, Azure or Google Cloud, and Docker.
Technical English
Comfort explaining model behaviour and its trade-offs directly to customer engineers.
Not required: You do not need a machine-learning research background or a particular degree; nothing here trains a foundation model. German is helpful, but English is the working requirement.
Tools
What you will work with.
- Python
- Hosted and open-weight LLMs
- RAG and embeddings
- MCP servers
- Evaluation sets
- Docker
- AWS, Azure or Google Cloud
The conversation
A clear path from hello to a decision.
Introduce yourself
Share something you have built or operated. Your message reaches Laura Deliu, who runs recruiting for the hub.
Talk engineering
A 30–60 minute first conversation about your work and what you want next. Ask for the salary range by email beforehand.
Work through a problem
A practical session together, inside working hours. No weekend take-home assignment.
Meet the hub
Talk with Noel Dinger about the company and what you would own.
Get a decision
A yes or a no, with a reason. After a technical session, feedback addresses the work discussed.
The practical questions.
Where is the role based, and what is the office schedule?
The role is hybrid and based in Cluj-Napoca. Home office is a normal part of the week, and the team meets at the Cluj office to work together. It is not a fully remote role, so you need to be able to get to the office. How the week splits between home and office is agreed in the first conversation.
Can I ask for the salary range before a call?
Yes. Ask in your application email and the range will be shared in the reply, before a technical session. The email preparation form includes an optional request for the range so you can raise it from the start.
Is this an AI research role?
No. Nothing in this job trains a foundation model or chases a benchmark. The work is choosing, prompting, connecting and measuring models that already exist, on data and systems the customer already runs. Fine-tuning comes up only after an evaluation set has shown that retrieval and prompting are not enough.
How is this different from the DevOps engineer role?
The DevOps engineer runs the platform the models are served on: Kubernetes, GPU nodes and the gateway that meters every call. This role owns how the model behaves on top of it: what it retrieves, which tools it may call, and whether a change made the answers better. The two work side by side on the same engagements.
Which models and stack will I actually use?
Python, mostly. The model is whatever the customer's data allows: frontier models in the customer's own cloud where they earn their cost, open-weight models on private infrastructure where the data cannot leave. Agents reach the customer's systems through MCP servers and connectors built for that customer.
What does the technical session involve?
A real problem of the kind the job contains, worked through with the people you would work with. There is no whiteboard algorithm round and no take-home task that expects a weekend of unpaid work. Expect to be asked what you would do when retrieval finds the right document at the wrong revision.
Your next move
Start with something you have built.
Tell us about a system, a difficult problem, or something you made better. A short introduction is enough to start the conversation.
Prepare an email. Review it. Send when you are ready.