Internship - Building an LLM-Based Assistant for Target Safety Assessments for 12 months

At Roche you can show up as yourself, embraced for the unique qualities you bring. Our culture encourages personal expression, open dialogue, and genuine connections, where you are valued, accepted and respected for who you are, allowing you to thrive both personally and professionally. This is how we aim to prevent, stop and cure diseases and ensure everyone has access to healthcare today and for generations to come. Join Roche, where every voice matters. The Position At Roche, Target Safety Assessments (TSAs) play a central role in early safety de-risking and translational decision-making. Over the yrs, Roche has generated a wealth of rich, expert-curated TSA documents—covering preclinical and clinical safety data, target biology, and risk mitigation strategies. However, accessing this information efficiently remains a challenge. This internship focuses on operationalizing a natural language chat interface for historical TSA documents, enabling toxicologists and project teams to query past assessments and extract relevant insights instantly. An NLP-based vectorization and agentic RAG pipeline has already been established internally. Your task will be to apply and refine this pipeline to process Roche proprietary TSA content and to develop a chatbot interface that provides semantically relevant, traceable answers to user queries. Join the Prediction Modelling team to collaborate with top experts, utilize cutting-edge bioinformatics and biostatistics tools, and gain invaluable toxicological insights in a dynamic scientific environment! The opportunity – Document Preparation & Vectorization: Apply the established Roche pipeline for document parsing,chunking, embedding, and indexing of historical TSA reports. Validate and troubleshoot document preprocessing (PDF/Word formats, metadata cleanup, section parsing). – Chatbot Development: Build or customize an interactive chat interface (e.g. using Streamlit, Gradio). Create rapid prototypes and production-ready tools. Ensure context retention, source traceability, and robust handling of safety-specific questions. – User-Centered Design: Work with toxicologists to understand key question types and develop sample QA templates (e.g., “What are the known liabilities of target X?” or “What safety concerns were reported for compound Y?”). – Evaluation & Optimization: Test the chatbot using both synthetic and real-world queries from safety scientists. Tune the embedding space, chunking strategies, and LLM prompts for toxicology relevance and performance. – Documentation & Knowledge Transfer: Prepare internal guides for deployment and future scalability. Present progress to relevant stakeholders (Translational Safety, Clinical Safety, Data & Analytics). Who You Are Enrolled in a Master’s/ PhD’s or final-year Bachelor’s program in computer science, data science, bioinformatics, or related fields – Strong experience with PyTorch, Hugging Face Transformers , and familiarity with LLM application frameworks such as LangGraph, Haystack, or LlamaIndex – Basic knowledge of LLM serving, and familiarity with LLM serving tools such as vLLM and SGLang – Working knowledge of embeddings, vector databases (e.g., FAISS, Chroma), and RAG concepts – Experience in implementing multi-agent systems by designing their core architecture and components from scratch, independent of pre-existing agentic frameworks – Basic knowledge of frontend and backend and interest in applied AI/NLP in the life sciences domain Non-EU/EFTA citizens must enclose a confirmation from the university that a compulsory internship is part of the training with their application documents. Start: October 2025 Duration: 12 Months Workload: 100% Who we are A healthier future drives us to innovate. Together, more than 100’000 employees across the globe are dedicated to advance science, ensuring everyone has access to healthcare today and for generations to come. Our efforts result in more than 26 million people treated with our medicines and over 30 billion tests conducted using our Diagnostics products. We empower each other to explore new possibilities, foster creativity, and keep our ambitions high, so we can deliver life-changing healthcare solutions that make a global impact. Let’s build a healthier future, together. Roche is an Equal Opportunity Employer.