The Long Road to AI Council

Anirudh Joshi — Thu, 22 Oct 2026 16:52:12 +0000

Hey everyone!

Myself Anirudh Joshi. I'm a 17-year-old, so-called developer—or, to be more precise, a vibe coder. I'm actually learning to code so I can finally unlock the official “Dev” tag.

For now, I build things and break them at the same time. I like breaking stuff even more, then figuring out how to fix it. Sometimes, though, I mess up so badly that it crashes me out and I end up deleting the entire project. SILLY Me,
Jokes apart, I've attended three hackathons so far, and I've won one of them! 🏆 I'm going to talk about the project I built that day, which won me a cash prize of ₹5,000. It might not sound like much, but it was my first win, and honestly, that feeling matters more than the money.

I've also built a web app called GoTradr, a virtual trading platform for people who want to learn about the stock market without risking real money.
The platform gives users ₹1 lakh in virtual money to analyze stocks, make investments, and track their profits or losses. It also has a news section that highlights stocks influenced by the day's news.

And the OP part of the app? Xenon AI. 🤖 It helps analyze stocks, explore their potential future performance, and make sense of the factors that could influence them.
You can check it out here: https://gotradr.base44.app/
I'm still learning, still building, and definitely still breaking things. More to come!

THE AI-COUNCIL

So basically, I was sitting in my room, thinking, “If I had ₹1 lakh, what would I do to grow it?”

So I asked Claude first:
"Hey, assume you have ₹1 lakh. What would you do to grow it? And don't tell me you'd do some freelancing and shit. Just give me some practical ideas."

Claude responded, but the answer didn't really match my expectations. So I asked ChatGPT. Its answer was a bit more satisfactory. I asked Gemini too, and got another decent answer.
But here's the problem: every AI model gave me multiple investment options, and I obviously didn't have enough money to try all of them. I was confused about which options to prioritize and how to narrow them down.
Then I had an idea: Why not build a tool where four AI models generate their own answers, and one chairman model analyzes their responses, synthesizes everything, and gives a final verdict?

And that's how AI Council was born.

Basically, this is how Newton discovered gravity when an apple fell on him. 🍎

The Tech Journey Begins

The First Mission: Finding Models That Actually Work

Throughout this journey, I used Codex with ChatGPT 5.6 TERRA at medium effort. I couldn't use high or ultra because, well, I'm broke.
The first mission was to find five AI models that were good at reasoning. I found a few candidates pretty quickly, but then came the real boss fight.

I selected 10 models and, with Codex, wrote a script to connect to OpenRouter's free models and test their reasoning capabilities.
And do you know what happened?
Out of 10 models, only TWO survived. DAMNNN.
I went through model after model, testing the free options on OpenRouter, and almost all of them gave up on me. At that point, even

I was giving up—just like I sometimes do with my studies.
Then, I had another idea. I asked ChatGPT, "Hey, suggest some alternatives to OpenRouter."
That's when I came across Hugging Face.
I already knew that Hugging Face offered access to AI models, but I wasn't aware of how to use its access tokens for inference. So I generated a token, configured it, and tested the models again.
KABOOM! 💥 The models I tested started running successfully.
Finally, we had a way forward.

The 12-Hour Hackathon

Once I'd figured out which models to use, it was time to build the actual application.
My vibe-coding skills kicked in HARD. I started putting AI Council together, working on both the frontend and backend while racing against the clock.
Meanwhile, my teammates took on what they considered the most difficult task of the entire 12-hour hackathon: making a PPT.
Jokes aside, we all had our parts to play, and somehow, I managed to get the application working in time for the judges.
The judges seemed impressed with what we'd built, and we ended up winning ₹5,000! 🏆
And one thing I genuinely appreciated was that my teammates later contributed their share of the prize money to me. That was a really nice gesture from them, and

I won't forget it.From struggling to find five working AI models to building an application under a 12-hour deadline, AI Council turned into something much bigger than the idea I started with.

The Hackathon Was Over, But the Problems Weren't

Soon after the hackathon, I took a break because I was completely burned out. And by a break, I mean a whole week of doing absolutely nothing related to AI Council.
After that, I got back to work. This time, I wanted to take AI Council to the next level by adding some cool features, improving the UI, and, most importantly, DEPLOYING IT!

I thought I'd already built the beast. All I had to do was deploy it, right? How difficult could that possibly be?
Well, destiny had other plans. 💀
I decided to call this version AI Council V2. The original V1 is available on my GitHub if you want to check it out.

I started adding Google and GitHub authentication to make signing in and creating accounts smoother. For the database, I went with Supabase's free tier because, as you already know, I'm broke.
I kept improving the loading screen, user experience, and mobile-friendly UI. Eventually, I finished what I thought was all the frontend and backend work.

Now came the testing phase.
I started with a simple prompt:
"Are we nearing the end of the AI bubble era?"
And guess what?
MY APP GOT STUCK ON THE LOADING SCREEN. 💀
I was sitting there thinking, "WHAT IS HAPPENING?"
Thankfully, my terminal was showing me what was going on in the backend. As I checked the logs, I discovered that all the models had failed, including the chairman model. The error was 402 — credits exhausted.

My first thought was that I'd run out of Hugging Face inference credits. Honestly, wouldn't that be the first thing you'd suspect too?
Since I was broke, I tried what I jokingly call my cheat trick: I created a fresh account and generated another access token.
I'd also built an option into the frontend that allowed users to update their Hugging Face token through their profile. The idea was that users could replace their old token with a new one without having to mess around with the backend themselves.

So I tested it.
FAILED.
I generated multiple tokens, tried again, and still got nowhere. At this point, I was getting seriously confused.
Then Codex came to the rescue, as usual. After investigating the problem, we discovered a mismatch between my V1 and V2 implementations that was affecting the connection between the frontend and backend. Turns out, the problem wasn't as simple as just generating another token.

And that brings us to where I am now.
AI Council V2 is built, but I'm still working toward getting it deployed properly. I haven't had much experience deploying a full application like this before, so there's still a lot for me to learn.
For now, I'm taking a step back, figuring things out, and learning what it actually takes to get an application from my local machine
onto the internet.

The goal is still the same: get AI Council out there and let people try it.
If you've deployed an application with a similar setup, I'd genuinely appreciate any advice. I'm still learning, and this is another part of the journey I'm figuring out as I go.
And who knows? Maybe the deployment story deserves its own article.