Ellen Drewes· 2026
Case studyData visualization · AI tools2025

From Three Weeks to Three Days

GraphWitch — How a broken chart workflow became an AI-powered visualization tool that locks the data and opens everything else.

San Francisco, CA · Remote

4 weeks · 2026

Shipped · Live at graphwitch.com

RoleFounder & Builder
ClientGraphWitch
DisciplineData visualization · AI tools
TIME SAVED3 weeks → 3 days

Overview

A major IGO client needed more than one hundred data visualizations for a report cycle, on a deadline that left no room for the way things had always been done.

The existing process was pure manual labor: designers reconstructed every chart by hand in Illustrator from Excel files, Word documents, PowerPoint decks, and pasted email fragments.

The goal going in was modest: cut the manual workflow by at least half. The result was GraphWitch, a tool that finished the entire pipeline in three days, revisions included, and then kept growing on the open web.

GraphWitch chart builder interface
Primary surfaceGraphWitch chart builder interface

Problem & approach

04 moves

AI models hallucinated values and changed data silently, with no flag or warning.

Each revision round compounded the risk, while formats, style guides, and source data kept shifting beneath the team.  Multiple stakeholders = multiple changes: in the data, the styles, the chart types and the formats. We needed a better solution.  Multiple stakeholders = multiple changes: in the data, the styles, the chart types and the formats. We needed a better solution. 

  1. 01

    Built an AI agent wrapped in strict guardrails: chart type, format, and style stayed open; the numbers stayed locked.

  2. 02

    Shipped a graph builder where users upload nearly any filetype, including images, and work with AI to extract data and construct visualizations.

  3. 03

    Added a style guide builder that generates an editable guide from a single image, with styleable variables and on-the-fly variations.

  4. 04

    Exposed the underlying data behind every chart for direct inspection and correction, so client edits applied at the source instead of by hand.

Workflow

Before & after

Before

  1. 01Collect data from disparate sources
  2. 02Check data for accuracy
  3. 03Build styleguide in Illustrator
  4. 04Build graph manually in Illustrator

After

GraphWitch interface extracting data from an uploaded image and generating a pie chart instantly
Graphwitch can take anything from an image to a spreadsheet, extract the data and build a graph instantly.

Iteration

Shipping live

Features to fix problems, in realtime

Each new snag in the project became a feature request, and I shipped them multiple times a day.

GraphWitch template picker with Dopamine, Grimoire, Seaglass, and Corporate palettes
Swap the entire visual system in a click — or extract a new template straight from a client's brand file.
GraphWitch chart type dropdown showing bar, stacked, line, pie, area, radar, and Venn options
Every chart type the team asked for, one dropdown away — no rebuild required.
GraphWitch extracted data table with editable rows and columns
The extracted data stays editable — tweak a value, add a row, and the chart updates instantly.
GraphWitch size preset menu with Small, Medium, Large, Slide, Full HD, Square, Instagram, Portrait, and Twitter/X options
Export presets tuned to the formats reports, decks, and socials actually ship in.
Extract a template from a file

Upload any image and GraphWitch pulls a full color template out of it — ready to apply to any chart in a click.

  1. 01
    GraphWitch template menu with an Extract new template from file option

    Start from any file — a screenshot, a photo, a client's brand deck.

  2. 02
    Interior photograph with a five-swatch palette extracted from the image

    GraphWitch pulls a full palette straight out of the image.

  3. 03
    Style name modal with the value Pinterest Room and a Create button

    Name it, save it, and it lives alongside your other templates.

  4. 04
    Bar chart rendered in the new Pinterest Room template with dusty pink, sage, tan, burgundy, and violet series

    Apply it to any chart — instantly on-brand, no manual restyling.

GraphWitch style inspector showing editable colors, title, description, and typography for an extracted template
Every element of the extracted styleguide is editable by the user.

Engagement rhythm

4 weeks · 2026

Problem to production, under 5 days

  1. 01Day 1

    Crisis

    Manual workflow breaks under deadline; off-the-shelf AI silently corrupts data.

  2. 02Day 3

    Prototype

    First guardrailed agent: locked data, open styling, one happy IGO team.

  3. 03Day 5

    Production

    Shipped graph builder, style guide editor, data panel, and multi-format exporter.

  4. 04Day 30

    Public launch

    Published on the open web; rebuilt interface around reliability, transparency, and ease.

Craft

Post-ship polish

Polishing the craft

Because I built for real-world needs on a deadline, the first version of the app was not exactly beautiful: it used default design patterns generated by Claude and Lovable. After the chart project shipped, I spent the time to fix major usability and visual design issues to make this usable for everyone, not just us.

Before — full app
Original GraphWitch UI with default Lovable-generated layout and generic sidebar
THE ORIGINAL GRAPHWITCH UI USED DEFAULT LOVABLE-GENERATED PATTERNS AND AN ALWAYS-EXPOSED STYLEGUIDE EDITOR (A SECONDARY USER ACTION)
After — full app
Refined GraphWitch UI with editorial typography, warm palette, and considered spacing
THE REFINED INTERFACE USES EDITORIAL TYPOGRAPHY, A WARM PALETTE, AND CONSIDERED SPACING.SECONDARY ACTIONS LIKE EDITING STYLEGUIDES ARE HIDDEN BEHIND A CLICK.
Naming a template
Default browser prompt asking for a template nameCustom Style name modal matched to the GraphWitch design system
THE ORIGINAL DESIGN HAD TECHNICAL ARTIFACTS AND SMALL BUGS THAT IMPACTED TRUST, ALONG WITH FONT SIZES THAT DIDN'T PASS ACCESSIBILITY. THE UPDATE FIXED THESE ISSUES. 
Data panel
Cramped data panel with harsh borders and tight spacing
Refined data panel with clear hierarchy, softer surfaces, and breathing room
THE DATA PANEL WAS HIDDEN BEHIND A TAB NEXT TO CHAT; MANY USERS MISSED IT. THE REDESIGN PLACED IT PROMINENTLY IN THE SAME VIEW. 
Chat surface
Original GraphWitch chat panel with default styling
Refined GraphWitch chat panel with editorial styling
THE CHAT SURFACE WAS TINY AND HARD TO SEE; THE USER'S PROMPTS WERE ALSO NOT SHOWING UP. THE REDESIGNED CHAT PANEL IS PROMINENT, LABELLED AND GROWS WITH THE CONVERSATION. 

Launch

Live product

Setting it loose: real users, no marketing

After the IGO project wrapped successfully, I wanted to see if the Graphwitch would be useful to other people too...so I published the site and made it free for all users. After just a few days, people started finding it – and using it. With no marketing or promotional efforts, people from all over the world were discovering this tool and using it, often for very long sessions. I'm watching closely, and gearing up to take it to the next level and turn this into a real, profitable product...stay tuned!

Analytics dashboard showing real GraphWitch visitors from around the world, with session duration and engagement metrics
REAL VISITOR ANALYTICS FROM THE FIRST MONTHS AFTER PUBLISHING — NO ADS, NO LAUNCH CAMPAIGN.

LIVE SITE OUTCOME — MEASURED

Instrumented
REAL USERS658
AVG TIME ON SITE8m
COUNTRIES14+
LONGEST SESSION42m
"
A chart tool that cannot be trusted with the data is not a chart tool.
— Working principle, GraphWitch project

Credits

Founder & Builder
Ellen Drewes
Platform
Lovable · Codex · Claude
Client
Major IGO (confidential)

Next

Quantcast
Data · Analytics platform

Quantcast analytics platform screen