In this special interview for Hackernoon’s “Startup Back” series, we live with Shashank Yadav, the founder and CEO of Fraction AI, a platform that strengthens users’ their own AI models. With a Background and Goldman Sachs and Core Machine Learning Teams in Goldman Sachs and Microsoft, Shashank gives information about his journey, AI scaling difficulties and how AI’s biggest bottleneck-safety, high-quality data. Take a look at how AI disrupts the AI view and how it democratizes AI ownership.
ISHAN PANDEY: Hi Shashank, a pleasure to invite you to our “Behind Start” series. Please tell us about yourself and tell us what inspired you to find the Fraction AI?
Shank yadav: Hey Ishan, it’s great to be here. Ben Shahank is the founder of Fraction AI. My past is in AI. I studied computer science at IIt Delhi, focusing on AI research. After that, I worked at the Core ML team in Goldman Sachs, then I made an early start as AI researcher and then moved to a hedge fund that applied AI to quantitative trade.
The problem I continued to run was that AI became the center. A few companies checked the most powerful models and everyone was stuck using ready -made versions that were not adapted according to their needs. But AI is not the only body that fits everyone. A lawyer needs a different model from a merchant or developer. The best artificial intelligence specialized, but his education was very expensive or very complex.
That’s why I started the Fraction AI. A platform where everyone can have their own AI models and educate. Users form AI intermediaries competing in sessions. Each representative pays a small entrance fee, creates the best output for a task and is evaluated by a LLM. The winners win awards and their models develop according to their best outputs. Over time, users create highly specialized AIs that continue to be better.
Instead of relying on several large models, we create an ecosystem in which thousands of smaller, specialized models compete, learn and grow. AI shouldn’t just have anything you use. There must be something you have and develop. That’s what we do.
Ishan Pandey: You worked in Core ML teams at Microsoft and Goldman Sachs. How did these experiences shape your approach to Frazir Artificial Intelligence Creating?
Shank yadav: Yes, during the university, I did an internship at the Bing team at Microsoft, and I worked on machine learning for the search ranking. This was my first real exposure to large -scale AI systems. The search is not just finding information, but to understand what users really want and to sort the results effectively. He taught me that AI is not just about smart models, but that they are about to work in the real world.
At Goldman Sachs, I was in the core ML team that created models for financial estimates. In finance, even small improvements are important and models are constantly tested under real world conditions where errors are costly. This experience taught me how to build reliable, adaptable and developing AI over time instead of performing well in a controlled environment.
Then, in a risk protection fund, I worked on AI for quant trade. I saw how strong competition could be. Models that constantly adapt and learn from competing strategies tend to perform better than static remains.
All this shaped fraction AI. Instead of creating an excellent AI, we have created a system in which AI agents compete, learned and developed according to the feedback in the real world. The best artificial intelligence is not designed alone – it develops by constantly testing against others. This is the idea behind the AI faction.
Ishan Pandey: You said the largest bottleneck of the AI industry is not power or programming calculation, but reliable data. Can you explain why the data is real restriction?
Shank yadav: Yes, I am based on this statement. The current AI models have already seen most of the internet. More calculating, if there is nothing new to learn, it will not help. The real difficulty is to obtain fresh, high quality data. Deepseek understood this and trained a model using pure reinforcement learning instead of traditional data clusters. They noticed that you can’t make fine -tuning on the same old data, you need a system that produces new and useful information.
We take this idea even further with Fraction AI. Instead of relying on static data clusters, we allow AI agents to compete in real world tasks. The best outputs are tried, refined and used to develop new generation models. Decentralized and constantly developing. AI should belong to everyone, not just a few companies. The best way to do this is to create a system in which people produce new, high quality data and train and develop their own models. Instead of locking artificial intelligence, it continues to develop with the real world use.
Ishan Pandey: What are the biggest misconceptions of companies in AI scaling, and how does the Fraction AI address them?
Shashank Yadav: The biggest error is that scaling is about AI’s more calculating to larger models. This worked in the past, but we hit a wall, more parameter does not automatically mean better results. Currently real bottleneck data, not calculation. Another mistake is to think that AI is static. Many companies make a fine -tuning on a model and assume that it is “finished”. However, it is not like AI software, but to continue to learn from new data to stay relevant. If your AI is not constantly developing, it will be behind.
Fraction AI improves it by improving itself by artificial intelligence. Instead of educating a model once and hoping to work forever, we create a system in which AI agents are constantly competing, learned from their best outputs and develops in real time. Not only scaling models, but about scaling learning. The future of artificial intelligence is not about creating the biggest model. It is about creating systems that can grow on their own. That’s what we do.
ISHAN PANDEY: What were the biggest challenges you encountered while switching to your own AI company from working in big technology?
Shashank Yadav: The biggest challenge was slipped from solving technical problems to a real company to business. In Big Tech, you focus on creating a model, but as a founder, product, users, financing and making sure that what you create is actually important – you should think of everything.
I spent a lot of time to watch Y Combinator courses to understand how to build and scaled how to build an attempt. IIt Delhi has a great culture of entrepreneurship, so I had many people to look at who took the jump. This gave me confidence that it was possible. Being a member of Nailwal was also a game exchange. Sandeep Nailwal, the founding partner of Polygon, is one of the most respected men on the web3 and was incredibly valuable to take guidance. He still understands how to build it in a clear, non -decentralized way while scaling things.
The most difficult part of establishing a company is not technology, it finds your vision into a real thing, how people actually use it. Learning from the others who made it before made a big difference.
Ishan Pandey: Fraction AI focuses on building a supported AI ecosystem. Can you disrupt how your platform provides scalable, high quality data collection?
Shank yadav: Fraction AI was built on the idea that AI should improve itself with competition and real world use. Instead of relying on static data clusters, we create a system in which AI agents produce, refine and develop data on a scale. This is how it works: users each form AI intermediaries with their own system request and setting. These agents compete in the sessions they produce for a particular task. Their answers are scored by a LLM Judge and the best performance representatives win awards. This process is continuously repeated and creates a feedback cycle in which AI models develop over time.
But we do not only collect data – we make fine -tuning on models. The best outputs from these competitions feed back to the training process and help the development and specialization of agents. During multiple sessions, users can upgrade their models, making them more appropriate for their smarter tasks.
This creates a scalable system for high -quality data collection and model development. Instead of relying on pre -existing data clusters, AI agents produce fresh, relevant data that are reality in real time. The result is an ecosystem that AI is not static – it always learns, always develops.
Ishan Pandey: What do you recommend to AI initiatives that try to navigate in the balance between innovation, sales and financing?
Shashank Yadav: Key timing. In the first days, focus on innovation and sales – you need enough product to prove what people want it, but you need to start selling early. Don’t expect perfection. If you can’t pay someone for that, it probably won’t solve a real problem.
Even if you have a small demand evidence, collect money as soon as possible. You need to survive long enough to build something great. Many initiatives fail because they focus on the product too much without securing enough runway. At this point, do not focus too much for dilution, initiatives are already zero or a game.
After collecting donations, everything is about sales and continuous innovation. Continue to develop the product while increasing income. If you can continue to sell and continue to push the technology forward, you will be ahead.
In short: When healing the product, demand → Rapid increase → Prove -scale sales.
EXPENSED INTEREST EXPLANATION: This writer,
