Why I Built an Offline Windows App When Everyone Else Is Going to the Cloud
Posted by Opposite-Future-5797@reddit | programming | View on Reddit | 13 comments
cazzipropri@reddit
FOUR HUNDRED AND TWENTY MEGABYTES?
Dude, what the fuck are you putting in 420 MBytes?
xpdf for win64 is 11.1 MB.
Opposite-Future-5797@reddit (OP)
Most of the size is from the bundled ONNX models and offline detection resources. The app does everything locally (AI detection, OCR-related stuff, multilingual patterns, etc), so a lot of the weight comes from shipping those dependencies instead of relying on cloud APIs.
cazzipropri@reddit
Ok, but my friend in Christ, what is "AI detection", given it's not OCR because you are listing OCR explicitly? Did you train yourself a language model to recognize personal information? It definitely doesn't seem to be the case, because your medium write-up doesn't mention it.
Tell me the truth - are you shipping a distilled LLM?
Opposite-Future-5797@reddit (OP)
Not an LLM đ
Itâs mainly smaller NER-style ONNX models for entity detection (names, locations, orgs, etc), combined with regex/rule-based detection for structured data. OCR is just the text extraction layer before that.
cazzipropri@reddit
Ok, we finally got to the meat.
Is that a 340M NER-oriented bert from huggingface?
I guess everything takes transformers these days.
Opposite-Future-5797@reddit (OP)
Pretty much đ
Smaller transformer-based NER models converted to ONNX, plus the runtime/dependencies around them. Also doesnât help that the Store install size ends up much larger than the original MSIX package after processing/distribution.
feudalle@reddit
We have a client we built a .net program handles all their dashboards, custom reporting, pdf and excel export. It comes in at 23 megs. Tbh its bloated and could be optimized. But I remember having to build for 1.44mb floppies and 4mb of ram.
cazzipropri@reddit
It's not an offline app.
It's a desktop, native application.
Opposite-Future-5797@reddit (OP)
By âofflineâ I mostly mean it doesnât rely on cloud processing or upload documents anywhere.
And yeah, regex already handles a lot of the structured stuff. The AI part is more for messy real-world text, OCR weirdness, names/locations, that kind of thing where regex starts falling apart.
cazzipropri@reddit
My friend, 95% of the desktop apps are like that. I'm not being rude. Take any app that people use on their desktop: Photoshop, Premiere, Outlook, VLC Player, Evernote, Illustrator, QuickBooks, AutoCAD... All the office and productivity applications are like that. MS Office constantly pushes you to the cloud, but that's because they want to sell you OneDrive and CoPilot. But, in general, desktop apps are almost universally designed to work locally.
Opposite-Future-5797@reddit (OP)
Fair point â âfully local processingâ wouldâve been more accurate than just âoffline.â
The main thing I wanted to highlight is that the app makes zero network requests, for privacy and trust reasons.
feudalle@reddit
But but ai. I remember when everything had to tie into the internet 25 years ago. Even if it was a lan application that used a local db.
xampl9@reddit
And XML!