Tools: We Tried the Chat-on-Left, Output-on-Right Pattern for AI Figures. It Failed. Here's What Worked.

The Problem ## Failure #1: The Chat Pattern Doesn't Work for Visual Composition ## Failure #2: AI Output Needs to Be Editable ## Failure #3: Single LLM = Single Point of Failure ## What Surprised Us ## The Stack (for the curious) ## Numbers So Far ## Try It I'm Mert — I have a background in bioinformatics research and I recently launched an AI tool for creating scientific figures. I wanted to share some honest lessons about the UX decisions and failures we went through, because I think they apply to anyone building AI-powered creative tools. Researchers spend an absurd amount of time making figures. A colleague of mine spent three days in matplotlib just trying to match color palettes across six figures after a reviewer asked for revisions. Three days. For colors. The existing workflow looks like this: We built Plottie AI to fix this. But the first version was wrong. Every AI tool in 2025 had the same layout: chat on the left, result on the right. We copied it. The problem? Researchers don't create one figure. They create 8-24 panels that need to look consistent. With a chat interface, each figure is an isolated conversation. You can't see them side by side. You can't compare color palettes. You always end up in PowerPoint anyway. What we built instead: An infinite canvas. Think Figma, not ChatGPT. Multiple AI-generated figures live on the same surface. You can: Result: one beta tester's workflow went from 90 minutes → 15 minutes for a complete multi-panel figure. Our V1 treated AI-generated figures as final. Generate → export → done. Researchers hated it. AI gets you 80% there, but the last 20% matters: exact hex codes for Nature's style guide, specific font sizes for figure legends, precise axis label formatting. "Close enough" doesn't work in academic publishing. Fix: We rebuilt everything to output editable SVGs. Every element (text, axes, legends, colors) is adjustable after generation. We also added 20+ one-click color palettes matching specific journals (Nature, Science, Cell, Lancet). V1 used one AI provider. When they rate-limited us or went down, our whole service died. On paper deadlines, that's unacceptable. Fix: Multi-LLM architecture — Claude, Gemini, and GPT with task-aware routing. Data plots go through a code sandbox (E2B), diagrams go through an Excalidraw-based pipeline. If one provider is slow, requests route to another. The user never notices. We built for data plots (bar charts, scatter plots, heatmaps). But the biggest surprise was demand for diagrams: flowcharts, CONSORT diagrams, pathway diagrams, scientific illustrations. Researchers don't just plot data — they explain processes. So we integrated a full diagram editor (Excalidraw) next to the data plotting engine. You can have a volcano plot and a CONSORT flowchart on the same canvas. If you want to play with it: ai.plottie.art — free, no card required. If you're building AI creative tools and ran into similar UX challenges (chat vs. canvas, single vs. multi-model, editable vs. static output), I'd love to compare notes in the comments. I'm Mert, building Plottie — an AI platform for scientific figures. Previously a bioinformatics researcher who was really, really tired of matplotlib. Templates let you quickly answer FAQs or store snippets for re-use. Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse - Write Python/R code to generate each plot individually - Export each as PNG - Open PowerPoint or Illustrator - Manually arrange panels A, B, C, D - Realize the fonts/colors don't match - Go back to code, tweak, re-export, re-arrange - Repeat until deadline - See all your figures at once - Drag them into "Frames" for multi-panel composition - Export the whole thing as one PNG/SVG/PDF - Swap color palettes across 20+ journal-specific presets - AI Engine: Python + FastAPI + multi-LLM routing + E2B sandbox - Frontend: Next.js 15 + Excalidraw + Konva (canvas) - Backend: Go + Gin + Typesense (search) + Cloudflare R2 (storage) - Auth: Supabase (shared via cookie across subdomains) - Deployment: Cloudflare Pages (frontend) + Fly.io (backend) + Docker (AI engine) - Launched: January 21, 2026 - Beta users: ~3,000 - Figures created: 3,000+ - Most popular: volcano plots, heatmaps, flowcharts - Free tier: 15 credits/day (enough for several figures)