KM Kanav Mehta Quantitative AI Research

Agent-Driven Research · Reporting

Reports an agent can assemble and a PM will actually read.

report-forge turns Markdown and local chart PNGs into publication-grade PDF, HTML, and DOCX, driven by agents over MCP, with quality gates that mechanically refuse to ship a bad exhibit.

Back to Projects GitHub

Typography

Sell-side polish, programmatic.

  • Nineteen templates: editorial layouts, portfolio light/dark in the gold/teal QuantFlow identity, Cyan Ledger light/dark, bespoke, plus nine domain briefs from earnings recaps to desk synthesis. Each is a real Typst layout, not a CSS skin.
  • Cover infographics on page one: KPI strip, conviction verdict band, key-point cards, and a fixed three-scenario strip, compacted so the body breaks cleanly to page two.
  • Spreadsheet treatment for showcase tables (banded headers, zebra rows, tabular numerals, right-aligned numbers), identical in HTML and PDF.

Discipline

Gates, not vibes.

  • Every figure must pass mechanical checks (resolution, aspect, legibility, provenance) or the build fails, and no blurry chart ships.
  • Native Exhibit numbering with cross-references, so "see Exhibit 12" survives reordering and regeneration.
  • Burned-in provenance: as-of dates and data stamps on every exhibit, carrying the research engine's audit discipline into print.

In production

The rendering half of QuantFlow.

  • Renders QuantFlow's flagship equity outlooks: 20–30 page chart-led reports (MSFT, GOOGL, META, NVDA) from one pipeline, each with a different brief.
  • Headless agent notebooks with persistent kernel sessions and a live editor sidecar; agents and humans co-edit the same live document.
  • One scaffold call per report: agents pick a template, place the charts, and the forge enforces the layout contract.
  • A gold/teal identity for exhibits across dark and light print themes; the reports and the site share one visual system.