THE THING WE BUILT:

DocCrusher

You know that moment when you drag a 47-page employee handbook into ChatGPT and watch your monthly token budget disappear? Like, gone. Faster than free pizza at a startup.

We’ve all been there. Everyone’s feeding documents into LLMs these days: company policies, project specs, that .PPT from ‘19 that somehow still matters.

But we’re doing it wrong. We’re sending bloated files full of formatting and digital noise that LLMs don’t really need.

But manually converting every document? Nobody has time for that.

So we built a thing.

DocCrusher in action!

How it works:

• Upload PDF, DOCX, PPTX, PNG, or JPG files - all common formats, even the big ones

• DocCrusher pre-processes your documents (because most model APIs won’t accept your giant PowerPoint)

• Get two outputs: human-readable summary and hyper-efficient LLM format

• Watch your 2ManyMB document collection become a 2KB text file with all the context intact

• Copy individual outputs or save everything as one complete file

• Feed results directly into any LLM, chatbot, or AI tool of choice

  
  

Notes from building:

1. Efficiency matters more as you scale. You wanted ChatGPT to generate an Ideal Customer Profile for your company. So you feed in all of your org charts, one-pagers, pitch decks, etc. And you asked it to generate a nice PDF. Cut to a week later: you feed all those documents again plus your new ICP PDF to ask a question. Then you do it again the next week. The idea is clearly working, so you start thinking maybe this should be its own app, something other teams can use. Maybe a lot of other teams! Suddenly, passing around 12MB of files per request isn’t so great. A tiny 2KB .txt file would be pretty nice, right?

2. Tokens add up. You already had to upgrade your account to feed in all of those PDFs before running out of tokens. But now you’ve got your new app and realized: Token savings aren’t just about cost — they prevent failure. Why do your requests keep timing out? Smaller payloads mean fewer timeouts, less data loss, more reliable AI workflows. When you’re building apps that depend on LLM responses, consistency matters more than perfection. It's the difference between your MVP app working and timing out.

3. Built-in prompt engineering maximizes API access. Anyone can call OpenAI’s API. But knowing how to chunk and structure data for optimal LLM consumption is where the money’s at. DocCrusher bakes in prompt engineering best practices so you don’t need a PhD in token optimization.

4. Show, don’t just tell. DocCrusher shows both readable and ultra-compressed output for every file. You see exactly how compression works - no black box processing. Compare before and after, understand what you’re getting.

Built with ❤️ by Township

You’re receiving this because you signed up for Township’s “We Built a Thing” newsletter. We'll only send these when we, well, build things.