Building Reproducible Financial AI Pipelines from Historical Documents and Live Data

Financial analysis increasingly combines two very different data sources: historical documents that should be reproducible and live information that changes constantly. IPO filings are a useful example. They can be long, multimodal, and difficult to process consistently, while questions about a company may also depend on current prices, filings, news, and macroeconomic data. In this talk, we build a reproducible financial AI pipeline around executive letters extracted from IPO filings. The workflow uses NVIDIA Nemotron OCR to recover text from document images, open language models served with vLLM to analyze what executives emphasize before their companies go public, and the You.com API to add current web and financial context. Flyte orchestrates the pipeline, tracks intermediate artifacts and lineage, caches expensive processing steps, and allows failed workflows to resume without repeating completed work. The goal is not simply to generate an answer, but to preserve how that answer was produced. We will look at what must be recorded when a workflow mixes static documents, OCR, model inference, and external data that may change between runs, and how to make the resulting analysis traceable, recoverable, and reproducible. Attendees will leave with a practical architecture for combining document AI, GPU inference, language models, and live external data in production workflows.
Who is this presentation for?
Data scientists, machine learning engineers, data engineers, and AI platform engineers building production workflows that combine unstructured documents, GPU inference, language models, and live external data. The talk should also be useful to practitioners working in finance or other regulated domains where lineage, recoverability, and reproducibility matter.
Prerequisite knowledge:
None
What you'll learn?
How to build reproducible AI pipelines that combine financial documents, OCR, open-model inference, and live external data while preserving lineage and recoverability.