Hello. I'm Ayush Kumar, a Full-stack Engineer and Designer. I went from a ₹2,000 freelance page to holding a patent and today I build a marketplace used by 11K+ people.

Currently, Technical Lead at Graminx. Before that, AI products at Growigh.

Always building, learning, exploring.

Build so far

I have loved learning and building things since childhood, whether it was experimenting with code, IoT projects or creating small projects just for fun, that curiosity turned into a passion, and over the years. Now I'm building more Stuff.

I believe that knowing how to sell and tell your story is as important as Building a great product. That's why I spend my time learning business strategy, understanding human psychology, and reading books on Design, Storytelling, and Startups.

I post on Twitter and write a blog on Medium.

Technical Lead · Graminx · 2026

Graminx

11,000+ users(as of Aug 2026)

Overview

Graminx is a location-based services marketplace for rural India, used by 11,000+ users. Customers book services and the platform matches each job to nearby, eligible workers, across farmer, worker, distributor and admin roles, with region-aware pricing and a multi-language interface.

How it works

Built a multi-stage booking-broadcast engine that matches service bookings to workers through a funnel of coverage-area and 25 km Haversine-distance filters, so each job reaches only eligible, nearby workers. Designed a location-based dynamic pricing system that resolves the best price through a Village to Region to District to State to global cascade (most-specific wins), letting admins set service prices at any geographic tier. Implemented a multi-role system (farmer / worker / distributor / admin) with scoped access, plus a multi-language UI backed by on-the-fly address translation and SMS/OTP auth.

Architecture

A Laravel monolith exposing a REST API to both the Blade and Alpine web app and the Android client, with Sanctum and SMS/OTP auth scoped across the four roles. It persists to MySQL, sends push notifications through Firebase Cloud Messaging, and resolves geography with the Google Maps and OpenStreetMap APIs. API keys and DB config were moved to environment and secret-based config, containerised with Docker behind Caddy, and shipped through a production deployment pipeline.

Tech stack

Frontend

Alpine.jsBootstrapTailwind CSSSass

Backend

PHPLaravel

Database

MySQL

Infra

DockerCaddy

Tools

Open Street MapGoogle MapsFirebase

Live ↗App ↗

Full-stack · Growigh · 2025

Ivaan

Amazon Startup Program

Overview

Ivaan is an AI meeting assistant, delivered as a SaaS. Its core value is simple: keep notes of your meetings and store the context for the next one. An AI agent joins your meeting, records and transcribes it, summarises it, and maintains the context of those notes across meetings, all accessible through the agent's website. The vision was to grow from this core into an all-in-one personal assistant, adding email management, calendar management and to-dos to help you hit your daily and weekly goals. Ivaan was selected for the Amazon startup program, but we eventually dropped it, as competition in this space grew very fast and it became hard to keep up.

How it works

A bot joins the meeting through a headless Chrome tab and captures the audio (a Chrome extension using tabCapture plus an offscreen document). The audio is transcribed with Whisper and run through speaker diarization to label who said what, then summarised. Notes and their context are stored so the assistant can recall them across future meetings. Work is queued and coordinated between services with BullMQ and Redis.

Architecture

A microservices system with separate services for the meeting bot (chrome-bot), transcription, and speaker diarization, plus a Chrome extension for tab capture, coordinated over a queue (BullMQ / Redis) and WebSockets. Runs on Docker Compose with Redis, PostgreSQL, MinIO (S3-compatible storage), and Prometheus + Grafana for monitoring.

Tech stack

Frontend

Next.jsTailwind CSSshadcn/uiTypeScript

Backend

HonoFastAPIPythonBullMQWebSocket

Database

RedisPostgreSQL

Whisperpyannote-audio

Infra

TurborepoDockerMinIOPrometheusGrafana

Tools

PuppeteerChrome Extension

Live ↗

Full-stack · Growigh · 2025

UI Analyzer

Live at uianalyzer.in

Overview

UI Analyser analyses the UI of any website. You enter a URL, and in the backend Playwright opens the site, captures a full-page screenshot and runs a Core Web Vitals audit. The screenshot is sent to an LLM with a prompt to critique the UI, and the results are shown back to the user.

How it works

Built a backend where Playwright loads user-submitted URLs, captures full-page screenshots, and runs Core Web Vitals audits. Implemented a scalable job system using BullMQ and Redis to queue the screenshot, LLM analysis, and audit work so requests are processed reliably under load.

Architecture

A NodeJS and Python hybrid backend. NodeJS runs the API, while Python workers handle the browser automation and image extraction. Everything is containerised with Docker.

Tech stack

Frontend

Next.jsReactTypeScriptTailwind CSSshadcn/ui

Backend

Node.jsExpressBullMQPythonFastAPI

Database

MongoDBRedis

LangChainGoogle GenAIOpenAI API

Infra

TurborepoDockerCaddy

Tools

ClerkStripePlaywright

Live ↗GitHub ↗

Backend · DK Tech (UK client) · 2024

LiveGame

100+ users

Overview

A live, real-time gaming platform where people play games. It has 6 user roles with role-based authentication, live data streaming to the frontend, and full user management. This project taught me a lot: writing complex real-time gaming logic, working with AWS (EC2, S3, CloudFront), Docker containerisation, and deploying.

How it works

The heart of this project was writing the complex real-time gaming logic and keeping live data in sync across all clients. I set up role-based authentication for 6 roles, streamed live third-party data to users over WebSockets, and handled all the user CRUD and transaction logic in the main backend.

Architecture

Split into separate services. A stateless Mirror API proxies a paid third-party feed for live cricket and casino data. A cron service polls the Mirror API and stores matches and series in MongoDB, flipping status to live. The WebSocket service pulls live scores and fancy odds from the Mirror API, reads match data from MongoDB, caches in Redis, and streams to the frontend. The main backend (Express + Socket.IO) handles game logic, auth, users and all CRUD, persisting to MongoDB and PostgreSQL (Prisma) with media on AWS S3. Deployed on AWS (EC2, S3, CloudFront) with Docker containerisation, shipped through GitHub Actions.

Tech stack

Frontend

Next.jsReactTypeScriptshadcn/uiTailwind CSSRedux ToolkitTanStack QueryReact Hook FormZod

Backend

Node.jsExpressSocket.IO

Database

MongoDBPostgreSQLRedisPrisma

Infra

AWSDockerGitHub Actions

Research · Amity University · 2022

Emersym

Patent · 3 hackathons · State-level recognition

Overview

EmerSym is an emergency detection and SOS alert system, built as a research project at Amity University Gwalior with my team. The app listens to environmental sound through the phone mic and uses a deep-learning model to classify it as emergency or non-emergency. On an emergency it automatically fires SOS alerts by SMS to the user's close relatives, the nearest police station and hospital, sharing the victim's location through the Google Maps API. It reached about 92% sound-classification accuracy. We presented it at DevFest, and it won 3 hackathons. The three of us also hold a patent for this work, titled "System and Method for Early Detection and Notification of Critical Situations Using Ambient Sound Analysis".

How it works

The Android app continuously records audio and sends it to a Flask backend. The server converts the sound into a model-compatible format, runs the trained deep-learning model (loaded from a pickle file, using MFCC audio features) to predict emergency vs non-emergency, and uses the haversine formula to find the closest police station or hospital to the user. If the prediction is an emergency, an alert is sent to the user's phone with a timer, and if the user does not cancel it, the SOS message is broadcast to relatives and nearby services along with the location.

Architecture

An Android client streams ambient sound to a Flask backend, which converts it to MFCC features and runs a pickled deep-learning model to classify emergency vs non-emergency. On an emergency it starts a timer on the phone; if uncancelled, the Google Maps API and Haversine lookup resolve the nearest police station and hospital, and an SOS is broadcast by SMS to relatives and those services with the location.

Tech stack

Backend

PythonFlask

TensorFlowscikit-learn

Mobile

KotlinJetpack Compose

Infra

Heroku

Books & Screens

Books

01 Before you startup
02 How to win friends and influence people
03 The Design of everyday things
04 Building a story brand
05 Exactly what to say
06 Never split the difference
07 Designing for Emotion
08 Microinteractions
09 Don't make me think
10 Laws of UX
11 Refactoring UI
12 Thinking with Type
13 User friendly
14 Universal Principles of Design

Screens

01 The Social Network
02 The Imitation Game
03 The Founder
04 Moneyball
05 Steve Jobs
06 The Big Short
07 Pirates of Silicon Valley
08 How to become a tyrant
09 Nike
10 The Billion Dollar Code
11 The man who knew infinity

Hello. I'm Ayush Kumar, a Full-stack Engineer and Designer. I went from a ₹2,000 freelance page to holding a patent and today I build a marketplace used by 11K+ people.

Build so far

Graminx

Ivaan

UI Analyzer

LiveGame

Emersym

Koor

Books & Screens