Headless deck unwrapping (DocSend / Pitch / Notion → PDF → summary)

Takes a deck-share URL (DocSend, Pitch, Gamma, Notion, Figma, Canva) with optional password, completes the auth gate using a residential-proxy headless browser, sniffs the internal GraphQL response for the direct /d/<id> href, hands that to a third-party PDF converter, OCRs any image-only pages with Tesseract, then asks Haiku for a structured JSON summary (company, founder, one-liner, vertical, stage, traction, team, links).

20+ decks per week get summarised with zero analyst time. Founder forwards a DocSend link → 60 seconds later the summary appears in the team's tracker.

1. Provision a Browserbase account; install Playwright + Chromium.
2. Detect URL shape: /view/s/<space> needs the browser step; /view/<id> can skip straight to the converter.
3. For space URLs: open in Browserbase, fill the email gate, listen for presentation/graphql responses, walk JSON for an href containing /d/.
4. Send the resolved href + viewer email + passcode to docsend2pdf.com/api/convert.
5. Save PDF to data/decks/<deal_uuid>/deck.pdf.
6. Extract text with pypdf. If under ~200 chars, fall back to pdf2image + pytesseract OCR at 180 DPI.
7. Run a Haiku summarisation prompt; emit structured JSON; attach to the deal record.

• DocSend's WAF actively blocks AWS / GCP / Azure IP ranges. Running the browser step from a normal cloud worker will fail silently — residential proxies are mandatory for that one step.