Rishav Sharma Swiss Re

GitHub LinkedIn Medium Email CV

Swiss Re · Feb 2024 → 2025 in review

Nebula, a retrieval assistant

How an assistant answers from a pile of documents instead of from memory, and what that changed for underwriters.

Python TypeScript RAG Semantic search Palantir AIP

This was built inside Swiss Re, so the system itself isn't public. What follows is the idea behind it and the outcome from my CV, nothing internal.

The problem

To price a risk, an underwriter reads about the company behind it: news, filings, financial reports. That reading took days per case.

The idea: retrieve, then answer

A language model on its own answers from what it absorbed in training, which is out of date and can't be checked. Retrieval-augmented generation (RAG) changes the order of things: first find the passages that matter, then let the model answer only from those passages, and cite them.

documents search top passages question model answer [1] [3]
Fig. 1 — Retrieve, then answer. The model only sees the passages the search found.
  1. Ingest. Documents are split into passages and turned into vectors, numbers that capture what a passage is about.
  2. Search. A question becomes a vector too; the closest passages are the relevant ones. This is semantic search: it matches meaning, not exact words.
  3. Answer. The model gets the question plus those passages, and writes an answer that points back to them.

What I built

PDF ingestion and the semantic search (Python for text extraction, TypeScript for the vector search), and an admin panel for prompts, evaluation and one-click reruns, all on Palantir Foundry. Underwriter research went from days to minutes.