Harshwardhan

001 — android

Utter

Wispr Flow–style dictation for Android, and on Android there wasn't one I could trust. So I built it for myself: hold a floating bubble, speak, and clean text appears in any app. It runs on your own Groq key and keeps a local word log with streaks and a weekly chart.

1.7^MB

release apk · ~0.4 s cold start

role: design, engineering
stack: kotlin · compose · canvas · whisper
model: bring your own key

read case study →

[utter][hold to talk]

“Let's move the review to Thursday, not Friday”

[words · 7 days]

002 — design · teaching

Spark

Spark is the OSS Club's intro to HTML, CSS and JavaScript for first-year students. I designed and built the benchmark site we show at the end, made from the same HTML and CSS they learn, to show students where the basics can take them.

400⁺

first-year students taught

role: design, build, teaching
stack: html · css · a little js

read case study →live site ↗

[spark 2026 · oss club]

003 — android · iqoo hackathon finalist

Smriti

A work log for people who don't work at a desk. Point the camera, say one line, and ask about it out loud later. The answer comes back with the original photo as evidence. OCR, speech and a Gemma 3 1B model all run on the phone.

0

network permissions · enforced at build time

role: team build · android
stack: kotlin · ml kit · gemma 3 1b
tested on: moto g05

read case study →demo video ↗github ↗

Smriti capturing a photo, extracting a record and answering a spoken question

[demo · moto g05 · 3× speed]

004 — web

Grounds

Google's AI answers cite their sources, but nobody checks whether those sources actually say what the answer claims. Grounds splits each answer into single claims and checks every one against the pages it cites, across markets.

148

searches · 8 markets · one recorded dossier