How I built 21st Fund? The complete technical roadmap is here..

Okay so quick context.

I’m an ex investment banker. Left my comfy IB job last year. Since then I’ve been vibecoding and learning how to actually build stuff.

Built a bunch of tools over the last 3 months, but one question kept bugging me:
will capital markets / VCs / gov programs even take vibecoded products seriously?

Initially I assumed no one would.
Then I thought… what do I even lose? Worst case I learn something.

So I shortlisted 3 products that I’d realistically pitch to VCs or apply for gov grants with.
And then I started looking for grants.

I thought this part would be easy.

Started with ChatGPT. Even deep research.
Bruh.

Expired links, grants that don’t exist, wrong deadlines.
Also my products were finance-heavy, so I wasn’t even sure if most grants applied to this niche.

Tried Grok, Gemini, Perplexity, etc.
Results were mostly the same. Perplexity was better, but still unreliable.

At that point I decided: screw it, let’s build this ourselves.

So I started by crawling the internet.

Initially used Puppeteer, then switched midway to Firecrawl because it scaled way better.
Pulled in everything I could find related to grants, credits, programs, incubators, banks, gov schemes.

Ended up with 100k+ grants.

Then wrote a bunch of data cleanup + analysis code to:

• normalize deadlines

• extract funding amounts

• clean eligibility

• dedupe overlapping programs

Filtered out expired ones (I know this is a tradeoff since some reopen later).
After that I was left with ~40k active grants / credit programs.

Next step: search.

Created a vector DB over Pinecone.
Used OpenAI Ada for embeddings.

I had like ~$500 of Pinecone credits left and built this entire thing in under 24 hours using agents.

At one point I thought about fine-tuning a custom model to reduce latency, but ditched it because:

• can’t keep adding new data easily

• retraining cost vs benefit wasn’t worth it

UI was next.

First time using Lovable and honestly… insane.
I gave it Groww (Indian finance app) as inspiration and it nailed the vibe.

Everything else is wrapped inside an Antigravity app.
Auth + DB via Supabase (still on free tier).

Models used for coding: Claude Opus + Gemini.

How the app works right now:

• user fills a questionnaire

• answers are compiled using GPT

• semantic search runs on the vector DB

• top 5–7 matching grants pulled

• for each grant:

  • real-time verification using Sonar search

  • eligibility + deadline rechecked

  • reasoning generated on why it fits or doesn’t

  • application score given (out of 100)

All grant data is stored dynamically and rendered in a clean UI.

End-to-end latency is ~40 seconds right now.
Not great, but I think it can be pushed down a lot.

Stack cost was under $100 to build.

If usage grows and egress becomes an issue, I’ll probably:

• move files to CF R2, or

• upgrade Supabase

End result:
user describes themselves once and gets custom grants / credit programs they can actually apply for.

Not selling anything.
Just sharing the dev experience and would love feedback or questions.