Case studies

The challenge

Physicians needed answers grounded in roughly 2,000 clinical guidelines and government healthcare documents, mostly PDFs with embedded charts and flowcharts. Retrieval had to stay accurate across a large knowledge base, handle questions that required calculations, and keep GPU processing cost under control.

What we built

A multimodal ingestion pipeline turned the documents, including embedded medical flowcharts, into structured knowledge using Marker, LlamaParse, Unstructured.io, and GPT-4 Vision. On top of it, an agentic RAG chatbot chose its own retrieval strategy, called tools for calculations, and queried a Neo4j knowledge graph for drug data, running independent steps in parallel on cost managed AWS GPU infrastructure.

The challenge

The signals that move markets hide in qualitative data: earnings calls, filings, transcripts, and podcasts that no analyst can read at scale. Shifts in management tone and anomalies in filings slip past until it is too late.

What we built

A proprietary analysis engine that ingests filings, earnings and conference calls, podcasts, and alternative data. Built on LLMs, vector databases, and financial APIs, it flags tone shifts, detects filing anomalies, and surfaces investment signals, backed by in house scraping platforms, RAG pipelines, and workflow automation.