Dev Dheeren BhamariaBCA · D Y Patil Deemed to be University · Navi Mumbai

Free, open-source tools that work with the Wi-Fi switched off.

Own your software.

I'm Dev, a BCA student. I rebuild everyday tools that usually cost money as free, open-source apps that run with no internet at all. Your files and your voice stay on your computer.

See the work Resume

2apps shipped,
both free

28PDF tools running
in your browser

0.82sto dictate 100 words
on a local GPU

0bytes uploaded
by either app

01Work

Apps that never ask for Wi-Fi.

Each one replaces something people normally pay for or upload their data to. Pick one to open it, or scroll through them in order.

01IHatePDF28 PDF tools that run inside your browser tabWeb · Windows
02WispernoPush-to-talk dictation on your own GPUWindows · CUDA
+Everything elseAll public repositories on GitHubgithub.com/Burthcer

IHatePDF

Project 01 of 02
Web app + Windows desktop app · v1.3.1

Twenty-eight PDF tools that run inside your browser tab. Nothing is uploaded, because there is no server.

Source code Download for Windows Full test report

Merge, split, compress, convert, sign, redact, protect, and edit the actual text inside a PDF. The file is read into memory, processed with WebAssembly, and handed back to you. It also ships as a Windows desktop app that saves results straight to your Downloads folder.

Tools: 28
Compression: 35.5 → 1.95 MB
Checks passing: 80
Network requests: 0

Edit text in placeDouble-click a paragraph and retype it. It reflows in the document's own font, and the old characters are removed from the file, not covered up.
Compress to a target sizeAsk for 2 MB and get it: a 35.5 MB scan came out at 1.95 MB, all on your machine.
Send to your phone, offlineShare files or whole folders over Wi-Fi, the PC's own hotspot, or Bluetooth.
Safe on low-memory PCsA memory fail-safe slows or stops a job before Windows runs out, instead of freezing.

01Inputyour diskYou pick or drop a file. The browser gives the page a File handle, and it never leaves the tab.
02Memorytab RAMThe PDF is read into an ArrayBuffer. Nothing is written to a server or to disk storage.
03Enginepdf-lib, pdf.js, WASMParsing, merging, compression and text edits run as JavaScript and WebAssembly inside the browser sandbox.
04OutputBlobThe result saves itself after a short countdown, or goes to your phone over the local network.
×Networkremote serversNONEEvery request was recorded during testing: none went outside the app.

Built with TypeScript · React · Vite · Tailwind · pdf-lib · pdf.js · Electron

Compressing a 300 dpi scanWindows 11 · Core Ultra 9 185H · 8 GB memory budget

10 MB2.9 s831 KB
50 MB9.5 s4.04 MB
150 MB26.6 s11.7 MB
500 MB91.4 s38.4 MB

One 35.5 MB scan, every level6 photos · output size

Light8.00 MB77%
Custom 2 MB1.95 MB95%
Balanced1.70 MB95%
Smallest341 KB99%

Startup: < 0.5 s
Memory at rest: 398 MB
Checks passing: 80
Requests out: 0

From the v1.3.1 test report: 17 + 34 automated checks and all 29 tools driven end to end in a browser, on Windows.

5,000-page text PDFspeak RAM · time

Job	v1.2.0	v1.3.0
Redact every page	8.3 GB · 454 s	1.2 GB · 779 s
Rotate	1.2 GB · 114 s	0.8 GB · 5 s
Organize (reverse)	1.9 GB · 30 s	0.8 GB · 5 s
Merge 5,000 + 5,000 pages	0.8 GB · 15 s	0.8 GB · 6 s
Edit a paragraph	1.1 GB · 12 s	0.9 GB · 7 s
PDF to Markdown	3.9 GB · 46 s	1.3 GB · 34 s
PDF to Word	1.2 GB · 16 s	1.3 GB · 22 s
PDF to Excel	1.3 GB · 39 s	1.3 GB · 51 s
PDF to JPG, 1,000 pages at 150 dpi	2.4 GB · 61 s	1.2 GB · 49 s
Split, crop, watermark, page numbers, compress, protect, PDF/A, sign	0.7 to 1.1 GB, 4 to 7 s	0.7 to 0.9 GB, 4 to 11 s

Conversions to PDFpeak RAM · time

Job	v1.2.0	v1.3.0
Excel to PDF, 100,000 rows (3,704 pages)	4.3 GB · 56 s	1.9 GB · 97 s
Images to PDF, 300 photos (367 MB)	3.1 GB · 15 s	1.4 GB · 21 s
Word (150 chapters), PowerPoint (150 slides), HTML (3,000 paragraphs)	0.8 to 1.5 GB	0.8 to 1.6 GB, 1 to 5 s slower

Every tool on every memory budget200 runs on Windows 11, installed app. Scans: 24 runs per budget. Student files: 28.

finishedslowed, then finishedstopped safely

624 MB budget (tight 4 GB PC), scans10 / 14 / 0
624 MB budget, student files19 / 9 / 0
1 GB budget (4 GB PC), scans15 / 9 / 0
1 GB budget, student files24 / 4 / 0
3 GB budget (6 GB PC), scans24 / 0 / 0
4 GB budget (8 GB PC), scans24 / 0 / 0
8 GB budget (16 GB PC), scans24 / 0 / 0

Compress, 1.3.0 vs 1.3.1 on the same PCWindows 11 · Core Ultra 9 185H

300 dpi scan	1 GB, v1.3.0	1 GB, v1.3.1	3 GB budget	4 GB budget	No limit
10 MB to 831 KB	stopped	2.9 s	2.9 s	2.9 s
50 MB to 4.04 MB	stopped	10.6 s	10.1 s	9.5 s
150 MB to 11.7 MB	stopped	31.3 s	26.0 s	26.6 s
500 MB to 38.4 MB	stopped	120.6 s	107.9 s	91.4 s	87.1 s

Windows busy with other programsfree memory shrinks as the job grows

Windows leaves free	Compress 150 MB	PDF to Word 150 MB	Watermark 500 MB
8,192 MB	finished, 33.7 s, 44/44 images	finished, 7.7 s	finished, 2.2 s
reserve + 762 MB	slowed, 38.4 s, 44/44 images	slowed, 29.7 s	finished, 1.9 s
reserve + 162 MB	stopped safely after 10.1 s	slowed, 39.0 s	slowed, 4.1 s

Last resorts

Case	Result
300-megapixel PNG to PDF at a 1 GB budget	Refused in 0.4 s; app usable afterwards
Windows down to 300 MB free during Compress 150 MB	Stopped in 2.4 s; app usable afterwards
A window stuck grabbing memory, 1 GB budget	Restarted in 1.5 s, notice shown
A job needing more than the budget (compress 1 GB at 0.7 GB)	Stopped after 2.9 s, memory back to normal, no restart

All 29 tools, driven through the browser like a person would29 / 29 passing, outputs reopened and checked

Tool	Input	Checked result
Merge	complex.pdf + test-multi.pdf	7 pages
Split	test-multi.pdf	selected range extracted
Organize	reverse order	4 pages, new order
Rotate	all pages right	rotations 90, 90, 90, 90
Crop	test-multi.pdf	crop box applied (555 × 735)
Compare	original vs edited copy	2 changes, +9 words, −18 words
Edit PDF	double-click a heading, retype	new text in file, old text gone
Sign / Fill forms	typed signature / “Zoë Test”	signature on page / form filled
Watermark, Page numbers, Redact	complex.pdf	3 pages each, redaction verified
Word, Excel, PowerPoint, HTML to PDF	.docx / .xlsx (120 rows) / .pptx / .html	2 / 6 / 3 / 4 pages
PDF to Word, Excel, PowerPoint, Markdown	complex.pdf	valid files; heading, table and list kept
Compress	35.5 MB scanned PDF	1.70 MB
Repair	complex.pdf cut off at 97%	all 3 pages recovered
PDF/A, Protect, Unlock	complex.pdf, password “secret123”	PDF/A metadata; encrypted, then decrypted
Auto-save	Page numbers, no clicks	saved itself after 5.3 s

Edge cases and privacy

Check	Result
RC4-40, RC4-128, AES-128, AES-256 password PDFs	wrong password refused, right one opens
Empty, garbage and renamed files	refused cleanly
Hindi watermark; Hindi and Chinese retyped in Edit PDF	correct characters, text extractable
Share to phone over Wi-Fi	QR, folders, ZIPs, wrong link refused: all pass
Network requests while using every tool	none outside the app

Windows checks, installed app

Check	Result
Startup	about 8 s on first launch after install, then under 0.5 s
App memory on the home screen	398 MB
App memory, 500 MB scan open in Rotate	1,424 MB, first 12 previews after 1.1 s
Watermark on a 500 MB scan at a 1 GB budget	stopped in 1.3.0; 2.6 s in 1.3.1
Network requests (every request recorded)	none outside the app

Desktop app (Electron)

Check	Result
Window ready after launch	0.47 s
Compress an 88.6 MB mixed-image file	3.06 MB, auto-saved to Downloads
Closing the window ends the process	41 ms
Hindi and Japanese output file names	saved exactly
Installer size	130 MB

Wisperno

Project 02 of 02
Windows desktop app · CUDA · Python

Push-to-talk dictation for Windows that runs on your own GPU. Your voice never leaves the machine.

Source code Install guide Full benchmarks

Hit the hotkey, speak, and clean text lands in whichever app has focus. Whisper large-v3-turbo transcribes on CUDA, then a local quantized LLM fixes filler words and punctuation. A free alternative to cloud dictation apps like Wispr Flow.

100 words: 0.82 s
Decode speed: 113.8 tok/s
Live mode: 13.5× real time
VRAM: ~3.5 GB

Two enginesInstant Direct Dictation with no LLM, or an on-demand Polish pass for grammar and filler words.
Live transcriptionA streaming mode for meetings and lectures, in a floating overlay with pause, save and discard.
TransformsPolish, Prompt Engineer, Bullet Points or Raw, plus your own custom prompts with their own hotkeys.
Zero audio egressNo network calls once the models are on disk. Speech is processed in local GPU memory.

01CapturesounddeviceWhile dictation is on, raw audio from the mic goes into a buffer in RAM.
02Speech to textfaster-whisperlarge-v3-turbo on CUDA. Live mode re-decodes the whole unconfirmed buffer every 0.6 s, so nothing gets dropped.
03Cleanupllama-cpp-pythonA quantized Qwen2.5-1.5B strips filler words and fixes punctuation. No API key.
04TypeOS hooksThe clean text is pasted into whichever app has focus. A small pill shows the current state.
×Networkcloud APIsNONENo audio streaming, no cloud transcription, no remote analytics.

Built with Python 3.12 · PySide6 (Qt) · CUDA · faster-whisper · llama-cpp-python

Four engine presets, measured on the dev machineRTX-class NVIDIA GPU · CUDA · Whisper loaded throughout

Turbo Flagship default113.8 tok/s0.82 s3,491 MB97.6%
Eco IQ4_XS103.3 tok/s0.89 s3,621 MB97.6%
Standard Q4_K_M91.7 tok/s1.04 s3,691 MB98.8%
Flagship Q8_063.1 tok/s1.48 s5,119 MB98.8%

Live mode RTF: 0.074
Re-decode every: 0.6 s
Engine upgrade: +20 to 30%
Audio out: 0 bytes

From BENCHMARKS.md, llama-cpp-python 0.3.26. Upgrading the engine from 0.2.90 made every preset 20 to 30% faster with the same VRAM.

Every preset, full numbersRTX 4070 Laptop GPU (8 GB) · Whisper loaded throughout · llama-cpp-python 0.3.26

Preset	VRAM	Process RAM	Decode	100 words	Fidelity
Turbo Flagship default	3,491 MB	2,311 MB	113.8 tok/s	0.82 s	97.6%
Eco (IQ4_XS)	3,621 MB	2,319 MB	103.3 tok/s	0.89 s	97.6%
Standard (Q4_K_M, 3B)	3,691 MB	2,559 MB	91.7 tok/s	1.04 s	98.8%
Flagship (Q8_0, 3B, full-precision KV)	5,119 MB	3,871 MB	63.1 tok/s	1.48 s	98.8%

Fidelity is measured on raw model output, before the app's own guardrail. In real use every transform goes through that guardrail, and all four presets pass the 5-prompt adversarial suite (prompts that try to make the model answer back instead of cleaning dictation) 100%. Every preset stays under the 5 GB system RAM budget with over 1.6 GB to spare.

Turbo Flagship decode speed after the engine upgradellama-cpp-python 0.2.90 to 0.3.4 to 0.3.26, version-bisected to find the newest release that doesn't crash on this GPU

beforeafter

Turbo Flagship89.3 tok/s113.8 tok/s

What changed

Measure	Result
Decode speed, every preset	20 to 30% faster
Output fidelity	identical
VRAM	unchanged (within 2 MB of the earlier clean baseline)
Cold start, time to first token after a mode switch	about 2× faster

Live transcription latencyfaster-whisper large-v3-turbo on CUDA · re-decodes every 0.6 s

Measure	Result	Target
Transcribe one 2.75 s window (word timestamps on)	~203 ms avg (199.8 to 209.9 ms over 5 runs)	under 200 ms
Same call on ~4.0 s of real synthesized speech	231 to 234 ms
Full hop (voice gate, transcribe, confirm) over 39 hops of 38 s speech	222 to 252 ms avg, max 570 to 928 ms	keep up with the 600 ms hop
Real-time factor on the 2.75 s window	~0.074 (about 13.5× real time)	under 1.0

The 600 ms hop is kept with over 2× margin even at the worst hop measured. An earlier sliding-window design was found by testing to silently drop words, so live mode now re-decodes the whole unconfirmed buffer every time.

Startup and real speech

Measure	Result
Whisper large-v3-turbo load (CUDA, int8_float16)	2.8 to 5.2 s across repeated runs; a fixed bug used to make it ~173 s
Local LLM load from its GGUF file	~1.1 to 1.2 s, all four presets
Real 9.75 s spoken sample, transcribed	0.945 s (RTF 0.097, about 10.3× real time), word for word
Peak during a live LLM transform	process CPU 92.4%, GPU compute 79%

How it comparesfrom the Wisperno README

	Wispr Flow	Local Whisper large-v3	Wisperno
Where audio is processed	Cloud (OpenAI + subprocessors)	On-device	On-device
Works with no internet	Never	Yes	Yes, after the first model download
VRAM floor	None (needs internet)	~10 GB	4 GB (falls back to CPU at 2 GB)
Cost per use	Subscription	None	None
Text cleanup	Cloud LLM	None built in	Local LLM, 4 presets

02Ideas

Three rules I build by.

If it runs on my computer, it should work without the internet. Everyday tools shouldn't be a subscription. And the safest upload is the one that never happens.

Ownership
Works without the internet.
Your documents and your voice belong on your machine. If a tool needs a connection to merge a PDF, the connection is there for the vendor, not for you.
Price
No subscription.
Build it once, own it forever. No monthly fee, no trial that runs out, and the source is open so anyone can check what it does.
Security
No upload, no leak.
A file that never leaves your laptop can't leak from someone else's server. Running offline removes whole categories of risk instead of patching around them.

03Stack

What I build with.

Everything here is in code I've actually shipped in the repos above.

01
TypeScriptIHatePDF
02
JavaScriptBrowser runtime
03
ReactApp UI
04
WebAssemblyIHatePDF engine
05
pdf-libPDF parsing & writing
06
Python 3.12Wisperno
07
CUDAWisperno GPU inference
08
PySide6 (Qt)Desktop UI
09
faster-whisperSpeech recognition
10
llama-cpp-pythonQuantized local LLM
11
GitVersion control
12
GitHubOpen source, MIT

Also

Offline / air-gapped testing
Reverse engineering
Threat modeling

04Education

Where I study.

Degree · Aug 2024 to 2029 (expected)Bachelor of Computer Applications

School of Humanities and Sciences. Coursework plus what I chase on my own: systems programming, operating-system internals, mobile apps, and taking software apart to see how it works.

Current student

Campus

D Y Patil Deemed to be UniversitySchool of Humanities and SciencesSector 4, CBD BelapurNavi Mumbai, Maharashtra 400614

05Contact

Write to me.

dev.bhamaria@gmail.com

Want to talk about offline-first apps, send a pull request to IHatePDF or Wisperno, or work on something together? My inbox is open, and I reply.

Resume github.com/Burthcer LinkedIn

Own your software.

IHatePDF

Wisperno

Works without the internet.

No subscription.

No upload, no leak.