Why Two IIT Engineers Turned Down $550K Jobs To Build A Startup
In a Nutshell
Two IIT engineers rejected $550K quant offers to join YC, pivoted from edtech to AI customer support agents after open-sourcing fine-tuned models, and won enterprise deals like DoorDash and Zepto by delivering working product with an 8-person team. They succeeded by validating real customer willingness to pay before building, focusing on product quality over sales, and fully committing without backup options. The core lesson: in AI, low build costs let founders test whether actual customers will pay for value delivered, rather than chasing ideas or market size analysis.
These notes were generated by AI and may contain inaccuracies.
Varun and his co-founder applied to YC with an edtech idea using LLMs. During the interview, Harj (HJ) told them the edtech idea would not work and to pick something else. They had prepped extensively for typical YC questions about ideas and TAM, but the interview focused entirely on their engineering background instead. Varun thought the interview went poorly and they would not get in.
"I was just panicked because I never prepared for anything."
Harj told them: "You guys are really good engineers. Just pick something else and work on it." Varun stated they would not have existed without Harj taking a bet on them.
Varun grew up in a small town in Andhra Pradesh where both parents were government teachers. He attended IIT in electrical engineering. His first two years were during COVID with limited studying. In his third year he began research on LLMs at Stanford, working on transformer models like BERT during the pre-ChatGPT period.
Varun received a $550K offer from a leading quant firm in New York and a PhD offer from Stanford. He and his co-founder decided to apply to YC instead. They got in three days before Varun was scheduled to join the quant firm. The co-founder was third-ranked at IIT with only one B grade across his entire academic career.
Varun's co-founder was described as very much a "book guy" who excelled academically. Varun described himself as more of a hacker who had poor grades but participated heavily in Kaggle competitions, earning approximately $50,000. He was eventually banned from some competitions for gaming the system. This experience helped him land a high-frequency trading job. Over time both founders became hacker-type builders.
After getting into YC, they pivoted from edtech after about a month in the batch. YC connected them with Coursera COO and other successful edtech founders, all of whom confirmed edtech was a bad idea. Both their B1 visas were rejected. They had experience with fine-tuning and read a research paper from a Databricks co-founder about LLM caching to reduce costs. At the time GPT-4 was very expensive. They decided fine-tuning smaller models might be better and open-sourced several models, topping Hugging Face benchmarks. This generated significant traction and led to a $4 million seed round.
Fine-tuning proved to be a difficult market. The primary reasons customers wanted fine-tuning were cost reduction and speed. A secondary use case was security for industries like insurance and healthcare, but these involved long sales cycles rather than engineering processes. After roughly a year they realized the two use cases growing fastest among their customers were customer support and coding.
They chose to focus on customer support based on actual customer usage patterns rather than market analysis. Zepto became their first customer for the customer support product after reaching out during Zepto's rapid scaling. The team did not know about competitors like Sierra or Intercom when they signed Zepto.
Varun stated they did not think much about competition. Their mentality was focused on whether customers were willing to pay and whether the product delivered value. After Zepto, they competed against a well-funded company for DoorDash's business. With a team of only eight people, they won against a company with 400 people. DoorDash became a major reference customer. Varun noted this showed there was arbitrage in building a great product rather than relying on sales teams.
Gary Tan introduced Varun to someone at DoorDash. Since DoorDash is itself a YC company, there was inherent trust. The team piloted for three months without any downtime and all metrics were strong. DoorDash was described as a very meritocratic company. Many subsequent large customers signed because of the DoorDash logo.
Current customers include one of the top three crypto exchanges in the US and multiple Fortune 500 companies. For AI agents, Varun stated the work fundamentally reduces to two things: policies or markdown files, and iteratively improving those files to move a business KPI. For support, the key metrics are resolution rate and CSAT. The same approach applies to compliance, IT, and HR use cases. Some of the largest US consumer companies are piloting the product for internal support and compliance.
Many people thought rejecting the $550K quant offer was a mistake. Varun and his co-founder simply wanted to see how high they could go. The co-founder is described as zero-motivated by money. Varun referenced Paul Graham's essay on wealth, noting it essentially comes down to starting a company and having equity in something big.
Varun's father was very upset about the decision. It caused a significant fight at home. Varun showed his parents YC videos and explained that even if the startup failed in a year or two, he could return to a job. Coming from a middle-class family created significant expectations, and even Varun internally questioned whether he was making a mistake.
The biggest mistake they made even after YC was working on multiple ideas that generated no revenue. Varun emphasized that ideas are not the scarce resource—ChatGPT can generate ten ideas instantly. The critical question is whether someone is willing to pay real money for the solution. Market size analysis is secondary. For any new product, they now require customer commitment and predicted payment amounts before building.
If a problem is important enough, people will pay either with money or time. Social networks are paid for with time. For B2B companies, if the problem is genuinely important, customers should be willing to pay money. Otherwise the problem is likely fake.
For customer-facing work, founders should stay close to customers. For research-heavy work in generative AI, San Francisco provides unmatched access to researchers and innovation, which is concentrated in the Bay Area.
Varun expressed strong confidence in the company's direction. The biggest bottleneck in enterprise AI deployments is the need for forward deployed engineers who sit with customers to configure systems. Giga ML is building an AI forward deployed engineer that will join Slack and Google Meet calls, take notes, and make policy changes automatically. This product is planned for launch soon.
One of the company's values is "automate, automate, automate." The broader mission is to automate all of the world's work. Sales teams use the system to pull transcripts from Gong and analyze what messaging works against specific competitors. The tool has turned many employees into builders who create their own automations.
Without coding agents, Varun estimates the engineering team would need to be six to seven times larger. The current team is relatively small but very strong. Benefits include faster shipping, reduced context switching, and the ability for individuals to own entire features rather than coordinating across many people.
The interview process requires candidates to write code, then removes AI tool access and asks them to modify the code without AI assistance. This tests understanding of how the code actually works. They look for "spikiness"—extraordinary ability in specific areas. Examples of spikiness in the founders include Varun receiving one of the highest quant offers and his co-founder ranking third at IIT and receiving India's highest-paying quant offer.
Many customers, including Zepto and DoorDash, did not care about traditional sales teams. The key is finding the right ICP (Ideal Customer Profile). Varun noted he initially believed sales was the most important function but was wrong. Looking at successful AI companies, product quality—not sales teams—drives success. No one uses Anthropic because of its sales team.
Anthropic and OpenAI do not pay sales people commissions, reflecting how little emphasis these companies place on traditional sales approaches. With AI, the product itself is the most important element—specifically how effectively it delivers substantial value to customers in a short amount of time. If this value delivery can be proven, everything else follows naturally.
After two to three years building the startup, having split time between San Francisco and Bangalore, several realizations have emerged about how the world works that were not obvious when getting started.
The biggest lesson is the value of getting started and actually trying to sell the product. Burning the boats is essential—committing fully without safety nets. This approach only becomes truly valuable when things get real. When the company was not working, both founders reflected on having rejected all job offers and questioned what they would do next. This situation forces action and compels founders to make things work.
Burning the boats does not necessarily mean rejecting every alternative permanently. If someone has a job, they can always get another job. The key principle is going at it fully and taking action, which holds much stronger value than maintaining backup options.
The cost of building things is so low. People should just build things and try to deliver as much value as they can to a very small set of customers and see if they can pay their money.
With AI specifically, the low cost of building makes experimentation accessible. The recommended approach is to build products, deliver value to a small set of customers, and test whether those customers are willing to pay.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
