Sofia, Bulgaria

Our client is a technological company at the forefront of the global medical research scene with real-world data. As a part of their growth, they need a Data Scientist whose main task would be to work on the logic for your assigned domain indications, with the Chief Data Scientist and the data science team close by. You design the algorithm, build the training set from real corrected cases, and check the result against a gold set before it ships. Getting the algorithm right is the job.

This is an AI-native environment so you build model-driven pipelines around your own workflows: agents do the first pass, you review and judge, and the system logs what got corrected. Every algorithm you ship leaves reusable automation behind it.

Responsibilities:

> Deliver analyses on real-world data.
> Design and implement domain-specific algorithms: the logic that turns clinical records for an indication into analysis-ready variables and cohort definitions.
> Grow into fine-tuning our self-hosted LLMs for clinical extraction: build training sets from corrected cases, run the fine-tunes with support, and evaluate against gold sets.
> Verify and validate the real-world data with model assistance: the model does the first pass, flags what it’s unsure about, and you judge the rest.
> Rebuild recurring delivery workflows as model-driven pipelines, so each month runs more automated than the last.
> Keep your work in proper repositories: versioned code, CI on GitLab, reviewable and repeatable.
> Work with both of their stacks: self-hosted vLLM for volume work, frontier models for planning and hard cases.
> Measure everything. Every automated step has an accuracy number; every analysis has a quality trail.
> Show the team how to run and extend what you build.

Requirements:

> AI-first thinking. Your default question is “can a model do this?” before you do it by hand even once.
> You use LLMs daily in your own work: coding agents, prompt pipelines, evals.
> You take data quality personally. In clinical data, a wrong value can distort an analysis a client acts on.
> You ship. A working flow this week beats a framework next quarter.
> You measure. An analysis without a quality check, or automation without an accuracy number, is not done.
> Solid Python for data work (pandas or similar).
> Strong SQL.
> Statistics fundamentals: distributions, testing, knowing when a result is noise.
> Hands-on LLM work: prompt pipelines, structured output, evaluation, and first experience with fine-tuning (LoRA or similar).
> Git as a habit, Docker and Linux without fear.

Brownie points:

> Survival analysis or other clinical and epidemiological methods.
> Airflow, dbt, or similar pipeline tooling.
> Experience with agent frameworks or multi-step LLM workflows.

Let's chat whenever at niki@cadabra.bg

(Recruitment License № 2709/ 17.01.2019)