The problem
Drafting tailored proposals — copy, layout, and product visuals — was the slowest, most error-prone step of the sales cycle.
What I built
- Built a multi-modal generation agent producing text + self-contained HTML (product cards, comparison tables, inline-SVG diagrams) and AI-generated images via the OpenAI Images API (gpt-image-2 → DALL·E-3 fallback).
- Closed the loop with an agent self-evaluation step — Playwright renders each HTML fragment headlessly (760×960 @2×) and a vision model re-reads it before presenting, catching layout errors automatically.
- Assembled each proposal section by section from reusable templates, so reps could swap in the right copy, products, and visuals per RFP.
- Exported finished proposals to InDesign (IDML) and Figma — print-ready, editable handoffs the design team can refine in their own tools.
System architecture
🎨
Architecture diagram — coming soon
Generation agent · HTML + image gen · screenshot→vision self-check · IDML/Figma export. (Upload diagram here.)
Demo
▶
Walkthrough video — coming soon
Prompt → generated proposal → self-eval → InDesign/Figma export.
Stack
OpenAI Images (gpt-image-2 / DALL·E-3)Vision modelPlaywrightHTML / inline-SVGInDesign IDMLFigma
Outcome
Higher win rate Through precise requirements coverage, partner-product inclusion, and faster drafting.