I tried to connect Varanasi's local shops to customers, gave up, then found a smaller problem worth solving

This is a post about a startup idea that did not work, and the smaller idea that came out of it.

The idea: local shops, delivery, before Blinkit arrives

In 2024 I wanted to connect local shops in Varanasi with the customers around them, and add delivery. The bet was simple: quick-commerce companies had not reached the city yet, and the kirana store around the corner already has the stock and the trust.

Then I started working out how it would actually run, and the same problems kept showing up.

1. Margin. Grocery margins are thin. Delivery cost, a rider's time and returns eat most of what is left, and a small city has fewer orders per kilometre than a metro.

2. The shopkeeper has to maintain an app. I was asking people who run a shop all day to also keep a catalog current on a phone. In my head it was a small ask. For them it is another job.

3. Prices change every day. Vegetables, flour, oil, milk. If the price in the app is stale, you either lose money or lose the customer's trust on the first order.

4. Investment. Riders, a warehouse or dark store, marketing, and the discounts you need to beat a habit. All of it up front, before you know the idea works.

What I found while researching: ONDC

While reading about how others solved the catalog and payments side, I found ONDC, India's Open Network for Digital Commerce. Instead of each platform signing up shops one by one, sellers publish once on an open protocol (Beckn) and any buyer app can find them. From my research, more than 1.2 lakh sellers were already connected to it.

That changed how I saw my own idea. A good part of what I wanted to build already existed, and it was being built by a lot more people than me.

October 2026: I gave up on the delivery startup

I had carried the idea since 2024, and by October 2026 I stopped. I could not see how I would beat the margin and investment problems as one person, and ONDC was already doing the part I had planned to build.

Then I saw Poppy

In October 2026, Sierra and Meta announced Poppy, the Personal Agent Protocol: an open standard so that a person's AI agent can work with businesses. Big consumer agents are adopting it.

I noticed something small. Agents will speak Poppy. Indian sellers already speak Beckn. Those two do not talk to each other, and a kirana owner is not going to implement a new US-designed protocol on his own.

That was the thing I could do without a warehouse, riders or a seller app. I do not need the shopkeeper to maintain anything: the catalog and prices live on ONDC already, and the seller's own quote is the price. I only needed to be the translator in the middle.

When I searched on 1 November 2026, I found Poppy SDKs and separate ONDC MCP servers, but nothing connecting Poppy to ONDC. That is a search result, not proof. If you know of related work, please tell me.

Haat

I called it Haat (हाट, the village market). It is an open-source gateway: a Poppy "Company" on one side, an ONDC buyer app on the other.

"do kilo aata aur 1 litre doodh"
  -> 2 x Whole Wheat Atta 1 kg, 1 x Toned Milk 1 L, Kashi Kirana Store, Lanka
  -> "Total ₹197, pay on delivery. Approve?"
  -> yes
  -> order KKS-1001 (placed once)

Each of my old problems got a different answer:

• Shopkeeper app and daily prices: not my problem anymore. Haat reads the seller's catalog from ONDC, and prices and terms come from the seller's own quote, never from what the agent claims.
• Investment: no warehouse, no riders, no inventory. It is software.
• Margin: I do not have an answer. Haat makes no money and I have not tested a business model. I am not going to pretend the old economics problem is solved.

Two design rules I would not bend

A language model never decides an approval. The person approves one exact revision of an order, through the protocol. The model never sees session tokens, keys, OTPs or full phone numbers.

Unknown is not failed. If an order may have gone through (a timeout, a connection reset, a 5xx with no rejection, a callback that never arrives), Haat answers 503 Retry-After instead of "failed". Telling someone "failed" when the order went through makes them retry, and then they get two bags of atta. An approved order runs at most once, even across restarts.

Hinglish without a model

My friends type lists like do kilo aata aur 1 litre doodh. I did not want an LLM filling a shopping basket, so the parser is rules plus a lexicon, and each step is testable: normalise digits from every Indian script, split on aur/और/commas, read quantity and unit, drop filler words, match through a phonetic key, and if nothing matches search the raw words so nothing is silently lost.

Two bugs the tests now pin: dropping a final vowel made makhana (fox nuts) equal makhan (butter), and "do" means both "two" and "give".

What is verified, and what is not

I ran npm run check on my current branch today: 176 tests, 176 passed, 0 failed, with typecheck and lint clean. Haat also passed the official Personal Agent Protocol validator over real HTTPS in an earlier run, 99 checks passed and 0 failed (results are in docs/conformance/ in the repo).

What it does not do:

• Everything runs on a local mock network of ONDC-style sellers. It has never touched a real seller, real money or ONDC's real registry.
• Going live needs a legal entity, whitelisting and ONDC's approval. Sign-in codes need a real SMS provider.
• Only Hindi was written by hand. Bhojpuri, Marathi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Punjabi and Odia were machine-drafted and need native speakers to review them.
• It is not affiliated with ONDC, Sierra, Meta or the Poppy project.

What I learned

1. A failed idea can still point at a real gap. I could not build the shop-to-customer business, but researching it showed me where two networks did not connect.
2. Check what already exists before building the infrastructure. ONDC had solved a large part of what I planned to build, and starting from it was better than starting from zero.
3. Do not let a model hold the pen on money. Approvals and prices come from the person and the seller.
4. Two years of an idea is not wasted if you stop on facts. Research is what showed me ONDC and, later, the gap next to it.

Try it and help

git clone https://github.com/techsaurabh-394/haat-gateway.git
cd haat-gateway
npm install
npm run demo

Haat is MIT licensed and PRs are welcome. Most contributions need no TypeScript:

• Review the machine-drafted language data (Tamil, Bengali, Marathi, and the rest)
• Teach the lexicon how your city spells pyaz
• Add a mock seller or a domain pack from your neighbourhood

If you build agents, I would like to hear how you handle the "unknown outcome" case. And if you run or know a kirana store, I would like to hear whether any of this matches how they actually work.

Live demo and docs | GitHub

I wrote about the technical side (Hinglish parsing and the approval gate) in more depth in my earlier post.

Built by Saurabh Pathak, AI engineer and full-stack developer from Varanasi.