I'm a backend and systems engineer in Dhaka, Bangladesh. Most of what I build starts with an annoyance I face daily, and ends when that annoyance is gone. I care about small binaries, low idle cost and no manual.

• EX-Expander: type a shortcode in any app and pick the expansion from a popup. No focus stealing, ~3 MB native build.
• LingoLens: snip a region of the screen and read the translation in place, over the original text.
• Left Right Audio Balancer: two volume sliders with live peak meters on a desktop card. Fixes single-channel Bluetooth headsets, ~390 KB native C++ with <1 MB RAM.

I have an AI-evaluation background across RLHF, tool-calling and multilingual projects, so I know how models fail and how to test for it.

Open to backend and systems roles, including AI evaluation work.

A system-wide text expander for Windows. Type a shortcode in any app, pick from the popup, and the expansion is inserted without stealing focus. You don't have to memorize your own shortcuts.

• ~3 MB native download, ~3% CPU while in active use
• Low-level keyboard hooks, so it works in browsers, editors and games
• GPU-rendered popups (Direct2D) that never take focus
• Per-app scoping, custom trigger characters, auto-start on login
• Clipboard-safe insertion, plus a WebView2 manager for adding and editing expansions

_{Type :hi, pick from the popup, inserted anywhere}

OCR screen translator. Press Alt+Shift+M, drag a box over any on-screen text, and the translation renders over the original.

• Persistent Flask + EasyOCR backend loads models once, so snips are instant across 20+ languages
• 55% faster OCR through image preprocessing
• Tiered translation fallback (Google, deep-translator, MyMemory) so a snip doesn't fail
• Dark PyQt5 control center for languages, colors, opacity and fonts

Per-ear volume controller for Windows that sits on your desktop as a widget. The setting exists in Control Panel, several dialogs deep, which makes it useless for anything you adjust more than once. This puts both channels on a desktop card.

• Linked mode holds the offset between the ears, so a balance you set survives the next volume change. Master scales both without breaking the ratio
• Live audio peak meters inside each slider show what is actively playing out of each channel
• Detects Bluetooth headsets that report a single channel, and fixes it by disabling AVRCP absolute volume with one click
• Rewritten in native C++ with Direct2D: ~390 KB standalone binary, ~0.8 MB working set RAM, 0% idle CPU
• Does not disappear on "Win + D" press, staying at the desktop layer instead of minimizing

_{Independent left/right sliders with live peak meters on Windows desktop}

• RWS TrainAI: 5,000+ multilingual evaluation and annotation tasks (Jan 2025 – Aug 2026)
• Outlier (Scale AI): 400+ assignments in LLM evaluation, adversarial prompting, RLHF and tool-calling, plus reviewing other contributors' work (Oct 2024 – Feb 2026)
• Completed RWS Linguistic AI certification training (machine translation, NMT, LLMs, MT evaluation)