I’m Dev Bhamaria. I build open-source tools that do serious work locally, keeping your files and voice on your machine.
Currently studying BCA while building privacy-minded desktop and web software.
Scroll for the workThe principle
Software should be powerful without becoming extractive.
Local processing is not a feature badge. It shapes the architecture: no upload queue, no account wall, and no dependency on a server being there tomorrow.
How I work ↓01 / Selected work
Three products, made end to end, from engineering constraints to the interface people touch.
01
Offline PDF toolkit
Web + Windows
A 28-tool PDF workspace that runs in the browser or as a Windows app. Files stay on the machine because there is no upload service behind it.
Built around browser-native processing, WebAssembly, and deliberately tested low-memory behavior. The project has public test evidence rather than a vague privacy claim.
02
Local dictation for Windows
GPU accelerated
Push to talk, dictate, and keep working. Speech recognition and cleanup run on a local GPU; the resulting text returns directly to the active app.
The real challenge was building an assistant that feels instant without quietly exporting sensitive voice data. The interface includes local history, a custom dictionary, reusable snippets, and controllable text transforms.
03
Offline voice assistant
Windows
Say “Jarvis” and it opens apps, clicks what is on screen, sets timers and reminders, and remembers things for you. Listening, understanding, voice and memory all stay on one Windows laptop.
Plain rules handle everyday commands in milliseconds, and a small local model is only used for messy speech. After acting, Jarvis checks the result before it says it is done, and risky actions need a spoken yes. From V1 to V4 it grew from voice commands to reading the screen by name.
02 / Working style
Start with the constraint.
Privacy, memory, hardware, and real-world failure modes are product decisions, not cleanup work after launch.
Interfaces should make capable systems feel calm. I care about the empty states, keyboard paths, and wording as much as the engine.
Where a claim matters, I try to test it, document it, and link the method. People deserve more than marketing language.
Based in
Navi Mumbai
03 / About
I’m a BCA student at D Y Patil Deemed to be University, building open-source tools alongside my coursework. My interests sit where product thinking meets systems work: local AI, thoughtful interfaces, browser capabilities, and software that lasts.
04 / Contact
Open to hiring opportunities, collaborations, product work, and interesting engineering problems. Write to me.
dev.bhamaria@gmail.com ↗Technical notes
Screenshots, engineering notes, and measured results live inside each project above.
IHatePDF is tested as a product, not just a demo. The published report covers tool flows, file outputs, privacy checks, and constrained-memory behavior.
| Evidence | Result |
|---|---|
| 35.5 MB scanned PDF at 2 MB target | 1.95 MB output, processed locally |
| Test report | 80 checks passing; no network requests |
| Stack | React, TypeScript, WebAssembly, pdf-lib, pdf.js, Electron |
Wisperno Pro keeps the entire dictation path on-device. The benchmark results below are from the documented local RTX 4070 Laptop GPU setup.
| Measure | Default Turbo Flagship preset |
|---|---|
| 100-word dictation | 0.85 s |
| Decode speed / cleanup accuracy | 110.2 tok/s / 98.3% |
| Local stack | Python, PySide6, CUDA, faster-whisper, llama.cpp |
Jarvis keeps the whole assistant on the laptop. Numbers below are lab measurements from the build laptop, an RTX 4070 Laptop GPU.
| Measure | Result |
|---|---|
| End of speech to action | about 650 ms |
| Click on a native window | 42 ms |
| Stack | Python, faster-whisper, llama.cpp, Kokoro, pywebview |
