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

chraem — Mon, 16 Feb 2026 18:14:11 +0000

Hey DEV community!

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.

The Problem

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:

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

We built Plottie AI to fix this. But the first version was wrong.

Failure #1: The Chat Pattern Doesn't Work for Visual Composition

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:

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

Result: one beta tester's workflow went from 90 minutes → 15 minutes for a complete multi-panel figure.

Failure #2: AI Output Needs to Be Editable

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).

Failure #3: Single LLM = Single Point of Failure

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.

What Surprised Us

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.

The Stack (for the curious)

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)

Numbers So Far

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)

Try It

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.

Learning Chinese as a Developer: A Minimal Practice Method That Works

chraem — Sat, 01 Nov 2025 21:23:09 +0000

I Built TypingMandarin — Learn Chinese by Typing What You Hear (Because Flashcards Weren’t Working)

Hey devs 👋

This is a small side project I’ve been building to solve a problem I kept running into while learning Chinese (and helping others learn it).

I could recognize words when reading or listening…
…but when I tried to say them, especially with the correct pinyin + tones, my brain would blank.

Flashcards helped with recognition.
But not recall.

So I started asking:

When does recall actually happen?

For me, the breakthrough came when I practiced typing what I heard instead of just reviewing cards. Typing forces:

Listening → accurate sound perception
Active recall → pulling pinyin from memory
Muscle memory → reinforcing tones through repetition

That led me to build:

👉 https://typingmandarin.com
A simple web app where you listen to short Chinese sentences and type what you hear.

No accounts required.
No gamified distractions.
Just listen → type → reinforce.

Why this works (memory-wise)

There’s a well-documented principle in cognitive psychology:

Active recall + feedback strengthens long-term memory more than passive review.

Typing what you hear triggers:

Input processing (listening)
Retrieval (recall)
Precision correction (tones, spelling)
Repetition (muscle memory)

It’s the same reason people who take notes by hand remember more than people who highlight PDFs.
Effort builds memory.

Who this is for

Beginners who want pinyin / tones to finally make sense
Intermediate learners who understand more than they can say
Heritage learners wanting to reconnect Chinese in daily life
Developers who just want one consistent, low-pressure practice habit

You can do 5 minutes a day and it still works.

What I’m building next

Personal review mode
Playback-speed control
Shadowing mode
Voice input (experimental)

If there’s anything you’d like to see — I’d genuinely love to hear it.

If you're learning Chinese (or have tried in the past)

What actually helped you make progress?
Was it:

Immersion?
Flashcards?
Conversation practice?
TV dramas?
Music + lyrics?
Something else entirely?

I’m collecting methods → testing them → turning them into small daily drills.

Would love your thoughts 🙏
Thanks for reading — and happy learning ✨

—
https://typingmandarin.com

DEV Community: chraem