You are browsing as a guest. Sign up (or log in) to start making projects!

Hbp100

By @Erox

  • 8 Devlogs
  • 26 Total hours

Imagine someone just got released form hospital with a bunch of medical documents , prescriptions and now he is unable to read those all(might be due to different native language or low english proficiency or even too much papers), so they wd like to rely on a llm . Same goes for insurance , business logs or even a dev like me but it is scary to thing abt all our private data is directly sent to the llm, i mean real humans will review the data for preparing better dataset, i felt uncomfortable so i build hbp100. Its a ai to save privacy from ai .It is a ml model built with thousands of examples and it just swaps private data with place holders like [email protected] with [EMAIL_1] , the llm gets enough context(and if using the placeholder kills context my ml is designed not to mask or mask based on confidence ) to work without using users real data and after response my engine again swaps the place holder with the real data and as it runs on the users device .

Source
Ship #2

@Erox on Hbp100

What started as a small python privacy project eventually and became a full Rust implementation with a Python interface.
The current version has

A memory safe rust crate published on crates , it has an ml tht was trained on more than 5k examples to eval if a detected word shld be masked or not based on the context , near by words and it has achieved over 90% accuracy in doin it , the ml is faster than human blink and max can do 22k iteration/s , for python high level guys i made pyo3 bridge , and it works on python flawlessly too but pyo3 overhead, yeah python guys arent performance maniac like me also i made the extractor of what shld be masked or not nothing and let the user put it via rs or json.

Limitations ->
It isnt always right(classic ml)

The extractors need regex and regex mighnt catch the target masking unit sometime

No ocr for documents

It doesnt have built in stt so we need external stt fir voice to text and then masking

How it works->
It decides mask/release a defined regex unit and places a placeholder instead of the value and saves the value to metadata , based on context , on top of a user prompt , then send the masked prompt to llm , llm generates response and then the engine swaps the placeholder with the value saved on metadata cache.

  • 3 devlogs
  • 6h
Try project → See source code →
Open comments for this post

@Erox on Hbp100

26m 41s logged

Finished the final retouching and now the website shld work too , the website isnt part of this proj in hackatime but it is indeed , also my remaining 6 hrs of last debugging shld recieve stardust, i got a mail abt tht , otherwise i had forgotten abt it lol

0
0
14
Open comments for this post

@Erox on Hbp100

3h 12m 58s logged

This time i removed all the extractors , and made a new impl to add config from json files or code in python and rust both , it took me a lot of time also i added a website built on hbp100 this time for easier review but dont forget me crate

0
0
35
Open comments for this post

@Erox on Hbp100

2h 9m 25s logged

This time i fixed the session issue and wrote THE README MYSELF ,took me a lot of time still worth it

0
0
9
Ship #1

@Erox on Hbp100

Hbp100

After all the debugging, rewrites, benchmarks, broken builds now Hbp100 v3.1.1 is now release d..

What started as a small python privacy project eventually and became a full Rust implementation with a Python interface.
The current version has:

  • A native Rust core
  • Python support through PyO3
  • Ml-powered contextual privacy decisions
  • Reversible placeholder masking
  • Session-based metadata
  • Extremely low processing latency
  • Releases on GitHub and crates.io

There are still edge cases, the extractor can still be improved, and there are plenty of things I want to build next like rayon and multi threaded performance for more speed . It is not like , i have less speed(the current speed is enough to make flash accept defeat) but why not , i use arch btw.

and now i can finally sleep at ease .

  • 5 devlogs
  • 20h
  • 15.55x multiplier
  • 314 Stardust
Try project → See source code →
Open comments for this post

@Erox on Hbp100

2h 2m 11s logged

Finally session based metadata added now i can sleep in ease also it is v3.1.1 both on crates and github

0
0
52
Open comments for this post

@Erox on Hbp100

6h 55m 36s logged

This is probably the most hardcore devlog in this whole session, my task was simply write debug , ask gpt why it didnt work , then deepseek wrote code still didnt work and thts how 7 hrs passed me adding session with persistent memory also with a lot of optimizations , also v3.1.0 is on cargo now , go ahead and check it up for me uWu.

0
0
12
Open comments for this post

@Erox on Hbp100

3h 17m 54s logged

This time it is a huge improvement , it looks like v3 to v3.0.1 but i added pyo3 and python api’s , i mean now even python can use my rust package with rust like speed + safety in python , also i fixed some small bugs in extractor , also added benchmarks (the speed is insane , see on the tagged screenshot ) also the ml is hella fast , in python the ml model only took 3x time than a whole iteration in rust with extractors and pyo3 overhead , also now i have got some more ideas to add in my hbp100 i think those will make it even better and polished..

0
0
29
Open comments for this post

@Erox on Hbp100

5h 6m 45s logged

today finally i shipped hbp100 v3 on crates “https://crates.io/crates/hbp100/3.0.0” , this time no python runtime full rust even ml in rust thx to lightgbm3 , now it is far faster than the python pkl one , it took me 3+hrs just to train the model and to convert my old python logic into rust , then i had to do the dual readme setup to prevent the bug i came to meet on my last project too , now its fine i made a release on github too , in the crates release there is no “test.rs” but in git release i added tht , the tagged screenshot is result of test

0
0
5
Open comments for this post

@Erox on Hbp100

2h 50m logged

actually when i started the project i didnt knew abt stardance , still after v2.2 i used it , v2.2 was pure python and really fast but still it didnt convince me , so i went to v3 rust and i tried rust and python hybrid it sucked i am unable to debug the mess i made so i am thinking to write the ml portion in rust too , i have a live website for v2.2 too , i am new to onnx so it might take some time debugging , still my v2.2 was pretty much successful , one can try tht on “https://github.com/Erox-02/humming-bird-v3/releases/tag/v2.0.0” , so next will be v3.0 and i intend to ship the live website as a other project

0
0
35

Delete project?

Are you sure you want to permanently delete this project? This action cannot be undone.

All devlogs, followers, and associated data will be removed.

Followers

Loading…