This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

In Maharashtra, a marriage biodata (परिचय पत्र) is still a printed page. A family collects the details: name, birth date and the exact birth time, education, job, salary, father and mother, brothers and sisters, which village the mama lives in, gotra, kuldaivat, and a photo. Then someone takes it all to the DTP shop. The operator retypes it into a template, charges ₹100–300, and the family's details and the photo stay on a shared shop PC, or travel back and forth on WhatsApp.

So I built Parichay (परिचय, "introduction", said "pa-ri-chay").

TL;DR. You type the details the way you'd say them, in Marathi, Hindi, English or a mix. Gemma, running on the phone, fills in a form. Code, not the model, turns "sakali 6:40" into सकाळी ६:४० वा., and every number on the page. You check it, pick a Marathi, English or bilingual layout, add a photo and print an A4 PDF. Nothing is uploaded: no account, no server, and after the first visit it works in airplane mode. A biodata costs ₹0 to make, which is what makes it a business for the shop that sells it.

What I Built

• Say it once. One text box. "Mulga. Naav Rohan Suresh Patil. Janma 12 March 1997, sakali 6:40, Pune. Uchi 5'9". BE Mechanical, COEP 2019…" Marathi, Hinglish and English all work, and "Try an example" fills in a sample so you can see it straight away.
• Gemma fills the form. About 25 fields: personal details, education and work, family, horoscope details, contact and expectations. It also proposes the Marathi spelling of every name.
• You check it. Every field is editable. Marathi name spellings, and any date, time or height the code couldn't read for certain, are marked check. Nothing is guessed.
• Blank means left out. Any field you leave empty doesn't appear on the biodata. Families differ a lot in what they put on one, and the form shouldn't decide that for them.
• Three layouts. Traditional Marathi (॥ श्री गणेशाय नमः ॥, a maroon double border, Devanagari digits), Modern English, and Bilingual with both scripts on every line. Add a photo and it sits in the corner.
• Print it. Print → Save as PDF gives an A4 page the shop can print or the family can send.

Why not just paste it into ChatGPT? You can, for one biodata. But the details go to a server you don't control, a chat model will happily "correct" a birth time or convert a date wrongly in a confident voice, and you still have to lay out a Marathi page yourself. Parichay keeps the details on the phone, never lets the model write a number, and ends with a print-ready page.

How It Makes Money

Two people earn from Parichay: the person making biodatas, and me.

Their side hustle. Families already pay someone to make a biodata. Today that's the DTP shop near the station, and the job is 15–20 minutes of retyping. With Parichay it's a two-minute check and a print. So it works for:

• DTP and print shops, which already sell biodatas. At ₹150 a biodata, even five a week is about ₹3,000 a month, and the time saved goes to other print jobs.
• The person everyone in the building asks for help with their phone. With Parichay on their phone, they can make biodatas for neighbours and relatives for ₹100 each. They don't need a shop, a PC or design software. A biodata costs them ₹0 to make, so the whole ₹100 is income.
• Vadhu-var suchak mandals (marriage bureaus), which make a biodata for every member and want them all to look the same.

How I earn. Parichay stays free and open source for any family that wants to make their own. What I sell is the version a business wants:

My costs are close to zero. Gemma runs on the customer's phone, so there's no API bill and no per-biodata cost. The site is a free Render static site. So every sale is profit, and the first sale covers the whole business.

It's built, and it works offline too. A shop pays me on WhatsApp/UPI and gets a licence key. They paste it in once, and the phone checks it with an Ed25519 signature, with no server involved. The shop's name on the biodata comes from inside the signed key, so one shop's key can't print another shop's name. A Bureau key carries an end date. The only things saved on the phone are the key and the shop's number, never a family's details. To see it, tap Try a demo Shop key or Try a demo Bureau key in the app.

₹499 pays for itself after four biodatas. At ₹150 each, the fourth biodata a shop makes has covered the plan, and every one after that carries the shop's name and number to another family.

The code is open, so anyone could remove the check and run their own copy. I'm fine with that. A shop isn't paying for a lock. It's paying for its name on the page, setup and someone to call.

Demo

Try it: parichay-uq3d.onrender.com. Open it in Chrome on an Android phone or a desktop with a recent GPU, tap Try an example, then Fill the form with Gemma. The first run downloads Gemma once (about 1.5 GB, use Wi-Fi). After that, turn on airplane mode and it still works. No WebGPU? The form still works by hand.

To see the paid plans, tap Try a demo Shop key or Try a demo Bureau key, pick the Royal layout, and the shop's name appears at the foot of the biodata. The 30-second film at the top walks through the whole thing.

Code

The project lives in parichay/. One static page, no backend, no dependencies and no build step: the browser loads plain ES modules, and the domain logic has 31 node:test tests.

How I Built It

Gemma reads, code decides

A biodata gets matched on its birth time and date, so a wrong one is worse than a blank one. So I split the work the same way every time:

• Gemma finds each value in the text and copies it as written into a JSON object. Its only creative job is proposing a Devanagari spelling for names, and those always come back marked check.
• Code turns what Gemma copied into dates, times and heights, and formats them for the page. If the code can't read something, the field stays empty and is flagged. It never guesses.
• Labels (जन्म तारीख, शिक्षण, वडिलांचे नाव, मामा, कुलदैवत…) are fixed strings in one file. The model never sees them, so it can't mistranslate a heading.

The parsers are where the Maharashtra-specific knowledge lives:

Phone numbers keep their English digits even on the Marathi layout, because people need to dial them.

Why Gemma 2 2B and not the smaller Gemma 3 1B

In my previous project, Gemma 3 1B under WebLLM started repeating itself on long prompts. WebLLM gives the whole model a single attention window, and Gemma 3 mixes short local windows with global ones, so a long prompt either overflows the local layers or gets forgotten. A biodata prompt is a field list plus a family's paragraph, so I used Gemma 2 2B, whose window matches its context. It's a bigger download, but it reads the whole prompt. The pick is a small pure function: f16 weights when the GPU supports them (half the memory), f32 otherwise.

The prompt goes in the user turn, because Gemma's chat template has no system role. The reply is parsed defensively: code fences, chatter before the JSON and trailing commas are all tolerated, and unknown keys are dropped.

The browser is the only thing that prints Marathi right

My first instinct was a JavaScript PDF library. They don't shape Devanagari: conjuncts like क्ष, ज्ञ and र्य come apart into loose letters. Browsers already shape Devanagari correctly, so the biodata is an HTML page set in Noto Serif Devanagari, laid out in A4 millimetres, and printed with @page { size: A4; margin: 0 }. "Save as PDF" in the print dialog makes the file.

On screen the same sheet has to fit a phone. Every size on the sheet is written in one CSS variable, --u, which equals one millimetre of A4. In the preview --u is 100cqw / 210 (a 210th of the preview's width), so the page scales to any screen; in print it's exactly 1mm.

Nothing to leak

The family's details live in a JavaScript object in memory and are never written to storage. The only things a phone keeps are a shop's licence key and its phone number, so they don't retype them. The photo becomes an object URL that never leaves the page. A service worker caches the app, the fonts and the WebLLM module; WebLLM keeps the Gemma weights in its own cache. A test checks that the service worker precaches every file the app needs, because one missing file silently breaks offline mode.

The bug I caught

Birth times are where Marathi and English collide. My first parser treated "pm" and "ratri" (night) the same way, and it also knew that "ratri 2:30" means 2:30 in the morning. So "2:30 pm" came out as 2:30 a.m., because the night rule fired for "pm" too. On a biodata that's the wrong horoscope. The fix was to split them: "pm" means afternoon, and only the Marathi night words get the after-midnight rule. There's a test for both now.

The rest

• Gemma: Gemma 2 2B, q4f16 (or q4f32), through WebLLM on WebGPU, loaded from a CDN at runtime.
• Fonts: Noto Serif Devanagari, Noto Sans Devanagari and Noto Serif from Google Fonts, cached for offline.
• Hosting: a Render static site from one render.yaml, which runs the tests as its build step.
• Design: maroon and saffron, the colours of a wedding card. Light and dark for the app; the sheet itself is always paper.

Why Does Open Innovation Matter?

Because a biodata is private. It's a young person's photo, exact birth time, salary and parents' jobs. The honest answer to "where does this go?" should be "nowhere", and with a hosted API it can't be. With open weights, Gemma runs on the phone and the details are never sent anywhere.

Because the business only works at ₹0 a biodata.

With a paid API, every biodata would cost the shop money paid to a company abroad, and I'd have to charge a monthly subscription to cover the tokens. With open weights there's nothing to meter, so I can sell the Shop plan once for ₹499 and step out of the way.

Because it works where the shop is. A small shop's internet is flaky. Once Gemma is downloaded, Parichay doesn't need it.

Because nobody can take it away, and anyone can change it. Gemma's weights, WebLLM, the Noto fonts and Parichay's code are all open. A bureau in Kolhapur can add its own layout; someone can add Gujarati or Kannada labels. No permission needed.

Where a closed model would do better: Marathi spelling. A large hosted model would transliterate names more reliably than a 2B model on a phone. That's why name spellings are always marked check and the person confirms them, and I think that's a fair price for keeping a family's details off a server. It also costs a 1.5 GB download and a phone with WebGPU, and the form still works by hand without one.

My First User

Not yet. I built Parichay in the last few hours of the Week 2 window, so I haven't put it in front of a shop. Next I’ll take it to a DTP shop in Pune with one question: would you pay ₹499 to have your name on every biodata you make? I'll add their answer here.

Prize Categories

• Best Use of Gemma: Gemma 2 2B runs entirely on the phone through WebLLM and turns free-form Marathi, Hindi and English into a structured biodata. It copies values and proposes Devanagari names; code handles every number and label, so the model can't put a wrong birth time on the page.
• Best Use of Render: the whole app is a Render static site deployed from one render.yaml, with the test suite as the build step, so a failing test never deploys. Live at parichay-uq3d.onrender.com.

---

Built solo in Pune, October 2026. I drafted this post with Claude Code and edited it myself.