In this exclusive interview for HackerNoon’s “Behind the Startup” series, we sit down with Shashank Yadav, the founder and CEO of Fraction AI, a platform that empowers users to train and own their own AI models. With a background in AI and experience in core machine learning teams at Goldman Sachs and Microsoft, Shashank shares his journey, insights on the challenges of scaling AI, and how Fraction AI is tackling the industry’s biggest bottleneck – reliable, high-quality data. Get an inside look into how Fraction AI is disrupting the AI landscape and democratizing AI ownership.
Ishan Pandey: Hi Shashank, it’s a pleasure to welcome you to our “Behind the Startup” series. Please tell us about yourself and what inspired you to found Fraction AI?
Shashank Yadav: Hey Ishan, great to be here. I’m Shashank, founder of Fraction AI. My background is in AI. I studied computer science at IIT Delhi with a focus on AI research. After that, I worked on the core ML team at Goldman Sachs, then joined an early-stage startup as an AI researcher, and later moved to a hedge fund applying AI to quantitative trading.
The problem I kept running into was that AI was becoming centralized. A few companies controlled the most powerful models, and everyone else was stuck using off-the-shelf versions that weren’t tailored to their needs. But AI isn’t one-size-fits-all. A lawyer needs a different model than a trader or a developer. The best AI is specialized, yet training your own was either too expensive or too complex.
That’s why I started Fraction AI. It’s a platform where anyone can own and train their own AI models. Users create AI agents that compete in sessions. Each agent pays a small entry fee, generates the best possible output for a task, and gets judged by an LLM. Winners earn rewards, and their models improve based on their best outputs. Over time, users build highly specialized AIs that keep getting better.
Instead of relying on a few big models, we’re creating an ecosystem where thousands of smaller, specialized models compete, learn, and grow. AI shouldn’t just be something you use. It should be something you own and improve. That’s what we’re building.
Ishan Pandey: You’ve worked in core ML teams at Microsoft and Goldman Sachs. How did those experiences shape your approach to building Fraction AI?
Shashank Yadav: Yeah, during college, I interned at Microsoft on the Bing team, working on machine learning for search ranking. That was my first real exposure to large-scale AI systems. Search isn’t just about finding information, but understanding what users really want and ranking results effectively. It taught me that AI isn’t just about smart models, it’s about making them work in the real world.
At Goldman Sachs, I was on the core ML team, building models for financial predictions. In finance, even small improvements matter, and models are constantly tested in real-world conditions where mistakes are costly. That experience taught me how to build AI that is reliable, adaptable, and improves over time instead of just performing well in a controlled setting.
, at a hedge fund, I worked on AI for quant trading. That’s where I saw how powerful competition can be. Models that continuously adapt and learn from competing strategies tend to perform better than those that stay static.
All of that shaped Fraction AI. Instead of building one perfect AI, we created a system where AI agents compete, learn, and improve based on real-world feedback. The best AI isn’t designed in isolation – it evolves by constantly testing itself against others. That’s the idea behind Fraction AI.
Ishan Pandey: You’ve said that the AI industry’s biggest bottleneck is reliable data, not computing power or programming. Can you elaborate on why data is the real constraint?
Shashank Yadav: Yeah I stand strongly by that statement. Current AI models have already seen most of the internet. More compute won’t help if there’s nothing new to learn from. The real challenge is getting fresh, high-quality data. DeepSeek figured this out and trained a model using pure reinforcement learning instead of traditional datasets. They realized you can’t just keep fine-tuning on the same old data, you need a system that generates new, useful information.
We’re taking that idea further with Fraction AI. Instead of relying on static datasets, we let AI agents compete in real-world tasks. The best outputs get judged, refined, and used to improve the next generation of models. It’s decentralized and constantly evolving. AI should belong to everyone, not just a few companies. The best way to make that happen is to create a system where people train and improve their own models by generating new, high-quality data. Instead of AI being locked away, it keeps evolving through real-world use.
Ishan Pandey: What are the biggest misconceptions companies have about scaling AI, and how does Fraction AI address them?
Shashank Yadav: The biggest misconception is that scaling AI is just about throwing more compute at bigger models. That worked in the past, but we’ve hit a wall, more parameters don’t automatically mean better results. The real bottleneck now is data, not compute. Another mistake is thinking AI is static. Many companies fine-tune a model once and assume it’s “done.” But AI isn’t like software, it needs to keep learning from new data to stay relevant. If your AI isn’t continuously improving, it’s falling behind.
Fraction AI fixes this by making AI self-improving. Instead of training a model once and hoping it works forever, we create a system where AI agents constantly compete, learn from their best outputs, and evolve in real time. It’s not just about scaling models, it’s about scaling learning. The future of AI isn’t about building the biggest model. It’s about creating systems that can grow on their own. That’s what we’re building.
Ishan Pandey: What were the biggest challenges you faced while transitioning from working in big tech to founding your own AI company?
Shashank Yadav: The biggest challenge was shifting from solving technical problems to running an actual company. In big tech, you focus on building models, but as a founder, you have to think about everything – product, users, funding, and making sure what you build actually matters.
I spent a lot of time watching Y Combinator courses to understand how to build and scale a startup. IIT Delhi has a huge entrepreneurship culture, so I had a lot of people to look up to who had already taken the leap. That gave me confidence that it was possible. Becoming a Nailwal Fellow was also a game-changer. Sandeep Nailwal, co-founder of Polygon, is one of the most respected guys in Web3, and getting his guidance was incredibly valuable. He understands how to build in an open, decentralized way while still making things work at scale.
The hardest part of starting a company isn’t the tech, it’s figuring out how to turn your vision into something real, something people actually use. Learning from others who’ve done it before made a huge difference.
Ishan Pandey: Fraction AI focuses on building a self-supported AI ecosystem. Can you break down how your platform enables scalable, high-quality data collection?
Shashank Yadav: Fraction AI is built around the idea that AI should improve itself through competition and real-world use. Instead of relying on static datasets, we create a system where AI agents generate, refine, and improve data at scale. Here’s how it works: Users create AI agents, each with its own system prompt and tuning. These agents compete in sessions where they generate outputs for a given task. Their responses are scored by an LLM judge, and the best-performing agents earn rewards. This process repeats continuously, creating a feedback loop where AI models improve over time.
But we don’t just collect data – we fine-tune the models too. The best outputs from these competitions are fed back into the training process, helping agents evolve and specialize. Over multiple sessions, users can upgrade their models, making them smarter and better suited to their specific tasks.
This creates a scalable system for high-quality data collection and model improvement. Instead of relying on pre-existing datasets, AI agents generate fresh, relevant data that’s validated in real-time. The result is an ecosystem where AI isn’t static – it’s always learning, always improving.
Ishan Pandey: What advice would you give to AI startups trying to navigate the balance between innovation, sales, and funding?
Shashank Yadav: The key is timing. In the early days, focus on innovation and sales at the same time – you need just enough product to prove people want it, but you also have to start selling early. Don’t wait for perfection. If you can’t get someone to pay for it, it’s probably not solving a real problem.
Once you have even a small proof of demand, raise funds as soon as possible. You need to survive long enough to build something great. A lot of startups fail because they focus too much on the product without securing enough runway. Don’t focus too much on dilution at this point, startups are a zero or one game anyways.
After fundraising, it becomes all about sales and continuous innovation. Keep improving the product while scaling up revenue. If you can keep selling and keep pushing the tech forward, you’ll stay ahead.
In short: Prove demand → Raise fast → Scale sales while improving the product.
Vested Interest Disclosure: This author is an independent contributor publishing via our
