Lahin
Zaman

Using React Native's New Architecture, a Node.js/Express proxy and C++ Nitro modules, I built an iOS app that treats a university dining hall as a first-class food database. The result: a campus meal logs as precisely as a scanned package, instead of being guessed at three times a day. It is on TestFlight with active beta testers.

• A custom biomechanical volume-tracking algorithm routes secondary muscle fatigue, which stopped the false overtraining alerts that make naive trackers unusable.
• Dynamic barcode resolution and bidirectional Apple HealthKit sync mean anything with a barcode logs in two taps and the data lands where the user already keeps it.
• EAS pipelines deliver TestFlight builds and over-the-air updates, so a fix reaches every beta tester the same day rather than the next release.

Using Whisper on Metal, Silero VAD and streaming sentence-boundary TTS, I built a macOS voice agent whose entire pipeline stays on the machine. The result: time-to-first-word fell 79%, from 18.7 seconds to 3.9, and the per-minute cloud speech bill went to zero.

• Streaming at sentence boundaries rather than waiting for a full response is what turned an 18.7-second pause into a 3.9-second one, which is the difference between a demo and something usable.
• Hash-verified auto-rollbacks let the agent edit its own code safely; I also patched a regex bypass that was leaking MCP tool access past the permission layer.
• A macOS Electron overlay with TLS token authentication and a fail-closed bind policy keeps it reachable across devices without opening it to the network.

Using Python, Flask, the Gmail API and Twilio, I built a watcher that polls a public course-availability API and fires the moment a seat opens. The result: unlimited sections tracked campus-wide from a single API call, alerts under 200ms, and zero authentication setup for the people using it.

• One API call covers the whole campus, so adding a section costs nothing and the alert still lands in under 200 milliseconds.
• Email and SMS run behind one operator-managed credential, which means a student signs up without creating an account or granting anything.
• A name and department matching algorithm resolves common-name collisions, so the instructor rating shown next to a section is the right professor's.

Using the Python standard library alone, I built a tool that ranks who to ask for a coffee chat by how likely they are to reply, and drafts the ask. The result: vague networking advice becomes an ordered list, without scraping anyone and without a single byte leaving the laptop.

• Scoring on two separate axes, fit and reachability, surfaces the warm contact a keyword search buries, because a shared school beats a matching job title.
• Standard library only means no installs and no accounts, so it runs on any machine with Python and nothing is trusted with your contacts.
• It reads your own LinkedIn data export rather than crawling the site, which keeps it inside LinkedIn's terms instead of risking the account it is meant to help.