Sign up
All challenges
Most candidates spend 6–12 hours; take what you need.

Document Intelligence Pipeline

Take a batch of company filings or concall PDFs and turn them into structured, queryable summaries.

The problem

We link thousands of results, concall transcripts and investor-presentation PDFs per company, but store only the URL and metadata — no extracted text, no summary. It is effectively a pile of links, not research.

The real constraint is cost and reliability at scale: a summarizer that works on 10 clean PDFs and falls over (or burns your LLM budget) on 4,000 scanned, multi-column, inconsistently-formatted filings is not a solution.

What to deliver

  • A pipeline that takes a batch of real filing/concall PDFs, extracts text (including handling scanned or poorly-structured documents), and produces a structured summary — key numbers, guidance, tone, red flags.
  • A concrete plan for controlling LLM cost at thousands-of-documents scale (batching, caching, cheaper models for triage, when to skip).
  • A storage design for the output that a downstream feature could actually query.
  • Notes on what you'd do differently at 10x the document volume.

What a good submission looks like

  • Treats "the PDF is a scanned image" or "the PDF has three columns" as expected inputs, not edge cases.
  • Has an opinion on cost, not just quality — a summarizer that costs more than the subscription it powers is not shippable.
  • Produces output structured enough that another engineer could build a feature on top of it without re-reading your code.

We're evaluating

Text extraction qualityLLM cost controlStructured output designHandling messy PDFs

How to submit

  1. Build your solution, using any tools or stack you prefer.
  2. Write up your thinking — decisions made, tradeoffs, what you'd do with more time.
  3. Zip your code + write-up, name it yourname-document-intelligence.zip.
  4. Email it to us — the button below pre-fills the subject and a template.