Medixy Health

A retrieval-augmented LLM that turns complex medical documents into answers that are cited and easy to act on.

Role

ML Engineer

Focus

NLP · LLM · RAG

Understanding the data

Less hallucination. Better answers.

Every answer is grounded in retrieved passages from trusted clinical guidelines.

Instead of a single long prompt, the pipeline splits retrieval, reranking and generation.

Data

System flow

A careful pipeline for a sensitive domain.

Documents are chunked, embedded and stored in a vector index; a cross-encoder reranks the top passages before generation.

The model was evaluated on a held-out set of real user questions for accuracy, citations and safety.

Explore

Frame the ML task, metrics and data constraints.

Prepare

Clean, label and split the data into solid datasets.

Train

Train strong baselines, then iterate on the model.

Ship

Evaluate, deploy and watch for drift in production.

A medical chatbot shouldn't become another risk to manage.

General-purpose LLMs tend to hallucinate dosages, invent sources and sound confident even when the evidence is weak.

The goal was to build an assistant that answers only from verified sources and says clearly when it does not know the answer.