WATWriting Analytics Toolkit: from designing a classroom app to building research software end‑to‑end

1. 1

   Where it started.

   A one-window form: tick categories of measures, pick a folder and one text, choose where the CSV goes, press Start. One essay type; results only as a CSV.

   ⤢

2. 2

   A web app that couldn't scale.

   The first ask was a version on LEI's servers, so I designed and built a web UI in Angular. But the NLP core was too heavy on memory and compute to serve many researchers at once. Three backend developers worked to slim it, but it couldn't be done in time, so the web UI was shelved.

   ⤢

3. 3

   The pivot I led.

   Building a Figma plugin had taught me that Figma is a web app running through Electron on the user's own machine.

   So instead of scaling the server, I pitched a desktop app, with the heavy work on the researcher's own computer.

   It was an economic call too: LEI's resources were limited. I learned the stack and built it, by hand at first, before agentic coding came into my work in late 2025.

4. 4

   Designed for a researcher's own machine.

   A desktop app needed a way to organize files and analyses, so researchers manage their work inside it as projects, corpora and analyses.

   ⤢

   • 
   • 

   • The wording is local: you load files instead of uploading them, and results are written to your computer, so you open the CSV or preview it instead of downloading it.

   • Measures are chosen in levels, so researchers know what they're picking from 3,000+.

   • An analysis runs with a processing status while you keep working elsewhere in the app.

   • And since an analysis is just a corpus and a few settings, any analysis can be duplicated, to replicate conditions and control variables more easily.

   WAT Researcher: the desktop app.

   ⤢

   • 
   • 
   • 
   • 
   • 

5. 5

   Back end and packaging.

   I wrote the Python and SQLite back end and packaged the app:

   1. 01PyInstaller bundles the NLP core
   2. 02Electron wraps it all into one app
   3. 03I did the Apple signing so macOS will open it

   The NLP core, and the automatic signing and hosting, are the engineering team's work.

   My team and I ran two informal usability studies inside LEI, and the postdocs who tested it went on to use it.

   WAT Researcher: back end and packaging.

   ⤢

   • 
   • 
   • 

6. 6

   Back on the web, on my UI.

   Shipping the desktop app first meant researchers weren't left waiting.

   Once we had the time and resources, the engineering team slimmed the core: the Windows setup went from about 1.2 GB to 700 MB, and it's still shrinking.

   the Windows setup

   about 1.2 GB→700 MB

   They adapted my desktop UI for the web, with limits of 10 analyses in any 24 hours, 10 texts each.

   It's in internal beta:

   1. quick checks on the web
   2. large corpora on the desktop

   The web version, on my desktop UI.

   ⤢

   • 
   • 

What sets WAT Researcher apart is the difference between evidence-based NLP measures and LLM-based analysis: it produces some of the inputs an LLM might use.