Abhijay

I pushed my first commit on 17 December 2025.

Below is what I have built since.

I build small apps in India and try to put them in front of the people they are for.

Worth opening

  1. Asli KamaiFor delivery riders
  2. Aravalli WatchFor people defending the hills
  3. EARTH1st place, GIC 2030 AI Challenge
Asli KamaiFor delivery riders

A delivery rider knows what the app paid them today. They usually don't know what they kept after petrol, the bike EMI and the recharge.

  1. 1Log a shiftType it, say it, or photograph the earnings screenPhone
  2. 2Read itA vision model reads the photo; a language model parses the speechModel
  3. 3Do the mathsPlain code works out what was kept, per hour and per slot. No AI touches a numberServer
  4. 4Explain itThe model words a lesson from those figures onlyModel
  5. 5Show the weekTake-home, coach, a weekly quiz and a savings goalPhone

Shifts are saved on the phone. After Google sign-in they are backed up to Postgres.

A rider logs a shift and sees the real take-home for the week. The model never does arithmetic: plain code computes every figure, and the coach may only explain those figures, in English or Hindi.

Built
August 2026 to now
With
React, TypeScript, Node, Postgres

Open Asli Kamai Read the code

Asli Kamai home screen: 4,475 rupees kept this week out of 5,970 earned, with petrol, EMI and recharge shown as parts of one bar.
Aravalli WatchFor people defending the hills

India's oldest mountain range is being carried away by the truckload. Residents film it and the complaints get dismissed. A satellite's timestamp does not.

  1. 1Pick a placeA pre-marked area or your own outline, then two yearsBrowser
  2. 2Fetch the passesDry-season Sentinel-2 imagery from Microsoft Planetary ComputerOpen data
  3. 3Measure the lossThe browser compares vegetation, pixel by pixel, between the two datesBrowser
  4. 4Ask the modelA serverless function holds the key and sends the measurements to a language modelServer
  5. 5Hand over evidenceA brief, a year-by-year chart and a complaint draft in English or HindiBrowser

Pick a place in the Aravallis and two years, and see what the ground lost. The tool can also scan the whole belt for the biggest changes and answer questions about the map.

Built
September 2026
With
TypeScript, MapLibre, Sentinel-2, Microsoft Planetary Computer

Open Aravalli Watch Read the code

Aravalli Watch opening screen: sourced figures on lost hills beside a map of the range near Faridabad.
EARTH1st place, GIC 2030 AI Challenge

A companion for rural Indian families that you talk to instead of type into.

  1. 1Speak or showAsk out loud, or photograph a medicine label or a loan paper
  2. 2Turn it into textDeepgram transcribes the speech; Tesseract reads the document
  3. 3Pick the toolkitHealth, education, farming, women's rights, or the open chatbot
  4. 4Write the answerA language model on Featherless replies in plain Hindi or English
  5. 5Say it backGoogle text-to-speech reads the answer aloud, so reading is never required

Four toolkits: health, education, farming and women's rights, plus an open chatbot, in Hindi or English. EARTH took first place at the GIC 2030 AI Challenge in April 2026. I built it with Dishank Sen, who wrote the backend.

Built
March to May 2026
With
React, Express, MongoDB, Deepgram, Tesseract.js, Google TTS

Open EARTH Read the code

EARTH opening screen: the word EARTH in white on black with a button labelled Tap to enter in English and Hindi.

The record

Each bar runs from a project's first commit to its last. The dates come from git log, not memory.

    Point at a bar to see what it is. Select it to open the project.

    If you ride for a delivery app, or you work on the Aravallis, tell me what is wrong with these tools.

    Message me on X See everything on GitHub

    Discord