Custom app #1: Custom translations

Fast translations are bad quality. Good translations are expensive. A solution!

Going beyond prompt engineering to creating custom applications felt like the next frontier (worry not, I will publish something on cracking prompt engineering soon). So here I picked a problem to validate if it’s even possible for someone without an engineering background to make something functional. Hype is hype and reality is reality ya know? Must figure out the difference.

Why not custom apps with Claude & ChatGPT?

I just haven’t cracked Claude or ChatGPT custom apps…yet. But also, it’s super important to me that I’m able to accomplish my vision for app functionality in the least amount of time so I’m not incentivized to spend more time trying than I already have.

The issue with Claude custom apps is that debugging felt like whack-a-mole and the time expenditure outweighed the potential benefits. Plus, the token limitations are annoying and I was already a paying subscriber. I was downgraded or hit my usage limits quickly which splashed cold water on my enthusiasm.

The issue with ChatGPT custom apps is that it won’t remember my custom instructions which makes it just a glorified “Project” (which if you haven’t tried, is actually pretty neat).

Not writing it off forever, but definitely writing it off for now lol.

Enter Base44

HB mentioned Base44 in group chat as an easy way to build apps with natural language with more general token limits and the conditions were perfect: I was already looking for something like that and I was 5 hours delayed at the airport.

So I entered in my prompt:

Hello world! My first prompt for AI-built web app

Which got things moving:

Base44’s initial response to my prompt

And we’re off to the races…right? Not quite. Meet the learning curve.

Here’s what went wrong:

1. After 10 back and forth dialogues, I realized that the system has a hard time with PDF and Excel uploads. Fine, easy enough to save and export the dictionary in .docx format from Google Docs. The original component the AI used for extracting word pairs couldn’t handle .doc or .docx so it reduced the input to .csv only.

   

   This went on for a while…

   1. The fix: I gave up. Exporting to .csv is not the end of the world. I lost my will to battle the AI here.

2. The AI used to extracting translation pairs had a seriously hard time with my existing document which had the English term bolded followed by its Japanese translation after a colon. This was annoying because I had to re-format the whole thing to put it into a table with column headers.

   1. The fix: I told ChatGPT to format it, boom done in seconds.

3. The AI kept referring to internet context vs my uploaded dictionary. This is a major problem because if internet context is good enough, I wouldn’t need to build this app in the first place.

   1. The fix: Explicitly tell the AI that you must ONLY refer to [whatever document you have as reference]

4. It had a hard time translating net new documents. There were two problems: one, it couldn’t fine my custom dictionary to use for translations (it was uploaded at this point) and, two, it kept generating a blank screen on the Japanese side.

   1. The fix: Explicitly tell the AI that the reference document has been uploaded. If you look in the admin tab, you can see the actual data component name that you can reference. In my case, the data component was called DictionaryEntry so I told the AI to reference that in it’s translations.

   2. The fix: For the blank screen issue, just debug with the AI. Take a screenshot of it, upload it to your chat and explain the problem of the blank screen on the translation side (don’t just say fix it, the AI doesn’t know what’s wrong by default).

      

      Being able to debug by uploading a screenshot is way cool

What delighted me:

1. Base44’s own capacity to fix bugs and errors! It explained to me what was wrong and assured me that it’s working on it. Very reassuring.

   

2. It inferred certain components without me telling it explicitly. Why yes, I did want it to be able to save translated documents and allow the user to not only bulk upload into the custom dictionary but also add in entries manually.

3. It protects the app automatically with authentication (Google login). This is configurable in the admin panel but nice touch because I didn’t need to worry about it.

4. Analytics and logs are provided explicitly so you can see who has accessed you app and track all the changes made.

5. Minor quality of life feature: Base44 automatically generates an app description and logo. This makes it look polished and ready to use.

My conclusion

The proof of concept works! I double and triple checked the translations, spot checking and checking on a sentence by sentence basis. It works reliably, other users can access it, and the functionality solves a real headache.

This is when I started getting over my skepticism and really got excited about the potential of custom apps, automations, and the impact AI can have on GTM. In the next blog, we’ll build a custom app that will save me about 4 hours of work PER WEEK so subscribe and stay tuned!

Wait can I try it?

Sorry y’all, this first one has proprietary data. I’ll be adding in data partition so any authenticated user can upload, manage, edit, and save their own dictionary and translated documents.