I'm a founder-engineer from India. I design the schema, write the API, ship the three frontends, run the servers, and increasingly write the tools that let AI agents operate the whole thing safely.
Now building — Selzix
Sell it. Rent it. Get paid for it. One workspace. Selzix gives Indian sellers and hosts their own store on their own link. They can sell products, rooms and stays, the payments go straight to them, and Selzix takes 0% in fees. It costs a flat ₹5,000/month, and the first 14 days are free with no card.
|
🚀 Start from what you already have |
🛍️ Sell products |
|
🏡 Rent out stays |
💸 Get paid, stay compliant |
|
🎨 Look like a real brand |
📈 Grow and bring buyers back |
🤖 Agent-native. Admin and seller MCP servers with scoped keys let AI agents build, redesign and check stores safely. A screenshot service lets them see their own work.
Under the hood
| ~224k lines of TypeScript, in one Turborepo monorepo | 1 engineer (me) |
| ~100 Postgres tables, typed end to end with Drizzle | 3 apps built with TanStack Start: storefront, seller dashboard and admin |
| Hono API on Bun, plus Redis, Meilisearch and background workers | Vitest, Playwright E2E and visual regression test suites |
Also running — @RingtoneRobot
A Telegram bot that has reached 1.5M+ users. Send it any song or movie name and it replies with ringtones in seconds. There are no commands to learn and no app to install. It has been running since 2021, and about 25K people still use it every month.
| 1.5M+ users reached | ~25K monthly users | Since 2021 | 4 rebuilds, ending up serverless |
|---|
What it took to keep a 1M+ user bot alive on a solo budget:
- Broadcasting to 1M+ users under Telegram's ~30 msg/s cap. A Cloudflare Queue consumer sends in paced 25 msg/s waves. It respects
429 retry_after, retries failed sends, and moves messages that keep failing to a dead-letter queue. - Crash-safe resume. Broadcast progress is checkpointed in D1, which is strongly consistent, rather than KV, so a restart continues from where it stopped instead of re-sending. Users who block the bot are pruned in chunks that stay under SQLite's limit on bound variables.
- Moving 1.19M users to serverless. I moved the bot from a Python + MongoDB server to Cloudflare Workers and batch-imported the whole user base into D1.
- One deployment, many bots. A single Worker serves several bot identities, each with its own webhook route and its own user table.
- Search that doesn't go down. Queries fall back across several ringtone sources, so a broken upstream never means an empty reply.
TypeScript Hono Cloudflare Workers D1 KV Queues · earlier versions: Python aiogram FastAPI MongoDB Redis
- Agents are users too. If an AI agent can't use a feature through a well-scoped tool, the feature isn't finished.
- One type system, end to end. The Postgres schema, API validation and UI forms all come from the same types.
- Screenshots before claims. Visual regression and E2E tests run before anything ships to real sellers.
- Own the whole path. From
docker compose upin production to the checkout button, I'd rather understand every layer than rent it. - Ship, then sharpen. Early users teach you more than a perfect roadmap does.
- How to Start a Home Tiffin or Meal-Plan Service in India (2026): Registration and Taking Weekly Orders
- How to Start a Cloud Kitchen From Home in India (2026): Licences, Order Channels and a Launch Order
- Can You Sell Homemade Food Online in India? FSSAI, GST and What Your Website Must Show (2026)
- Starting a cloud kitchen in Hyderabad: FSSAI, trade licence and steps (2026)
- Starting a cloud kitchen in Chennai: FSSAI, trade licence and steps (2026)
Selling on Instagram? Try Selzix free for 14 days. Building something in commerce or AI agents? Say hi.
There is no END to Learning.