Dev Dheeren Bhamaria

TypeScript, React, Vite, Tailwind, pdf-lib, pdf.js, WebAssembly, Electron

• Built 28 PDF tools (edit, merge, split, compress, convert, sign, redact, protect) that run fully client-side in the browser or as a Windows desktop app. No file is ever uploaded.
• Wrote an in-place text editor that retypes a paragraph in the document's own font and removes the original characters instead of covering them.
• Target-size compression: a 35.5 MB scan to 1.95 MB for a 2 MB target; a 500 MB scan compressed to 38.4 MB in 91 s on an 8 GB memory budget.
• Added a memory fail-safe for low-RAM PCs and offline phone sharing over Wi-Fi, hotspot or Bluetooth. 80 automated and end-to-end checks passing; released with a Windows installer.

Python 3.12, PySide6 (Qt), CUDA, faster-whisper, llama-cpp-python

• Push-to-talk dictation for Windows: transcribes locally with Whisper large-v3-turbo on CUDA and cleans the text with a quantized Qwen2.5-1.5B. No audio leaves the machine.
• 0.82 s end-to-end latency for 100 words at 113.8 tok/s decode, in about 3.5 GB of VRAM on the default preset.
• Live transcription mode re-decodes the full unconfirmed buffer every 0.6 s at a real-time factor of 0.074 (about 13.5x real time), replacing a sliding-window design that silently dropped content.
• Benchmarked four engine presets and upgraded llama-cpp-python from 0.2.90 to 0.3.26, making every preset 20 to 30% faster with no change in VRAM.