Why Two IIT Engineers Turned Down $550K Jobs To Build A Startup
Varun, co-founder of GigaML, shares his journey from IIT student to building an AI customer support company, having turned down a $550K quant firm job offer. The company pivoted multiple times before finding product-market fit in AI customer service agents, winning major clients like DoorDash and top crypto exchanges. Varun reflects on lessons around selling before building, the outsized leverage of coding agents, and the value of 'burning the boats' to force real commitment.
Summary
Varun, co-founder of GigaML, recounts his path from a small town in Andhra Pradesh, India, where his government-teacher parents pushed him toward engineering, to cracking IIT and eventually building an AI startup in San Francisco. During college, he funded himself by winning Kaggle competitions — earning roughly $50,000 — which also led to a job offer from a top New York quantitative trading firm at $550K. Simultaneously, he received a Stanford PhD offer. When ChatGPT launched in December 2022, Varun and his co-founder decided to apply to Y Combinator instead of taking either opportunity.
Their YC journey was rocky from the start. The interviewer, HJ (Harge), rejected their original edtech idea in the interview itself, telling them to pick something else given their strong LLM research backgrounds. Varun panicked, thinking they had failed, but YC admitted them anyway on the strength of their engineering credentials. After YC arranged calls with edtech veterans who confirmed the idea was weak, the team pivoted within a month to LLM fine-tuning — open-sourcing models, topping HuggingFace benchmarks, and raising a $4M seed round.
Fine-tuning proved to be a difficult market due to commoditization and the enterprise sales complexity it required, so the team pivoted again, this time guided by what their actual customers were successfully using the product for: customer support and coding. Zepto became their first customer service client, and the team subsequently won DoorDash's contract as a team of just eight people competing against a 400-person, well-funded rival. Varun attributes that win to YC network trust, consistent uptime during a three-month pilot, and DoorDash's meritocratic culture.
GigaML now works with major enterprises including top-three global telecom providers, a leading US crypto exchange, and Fortune 500s, targeting resolution rates of 90–95% compared to the traditional 10–15% deflection rates of IVR/chatbot systems. Varun describes the core insight driving their product: AI agent performance fundamentally reduces to the quality of policy markdown files and the ability to iteratively improve them to move KPIs like resolution rate and CSAT.
Looking ahead, Varun is building an 'AI forward deployed engineer' — an agent that joins Slack and Google Meet, takes notes, and automatically implements configuration changes, eliminating the bottleneck of human forward-deployed engineers in enterprise AI adoption. Internally, GigaML operates with a very small but high-quality engineering team, citing coding agents (specifically Claude/Claude Code) as enabling roughly 7x engineering leverage. Their hiring process includes vibe-coding interviews followed by removal of AI access to verify genuine code comprehension.
In closing advice, Varun emphasizes that the most important validation signal is whether someone will pay real money for a solution, that product beats sales in the AI era, and that 'burning the boats' — truly committing with no safety net — is what forces founders to build something real.
Key Insights
- Varun claims GigaML's fine-tuning pivot was driven not by market research but by observing that only two use cases among their actual customers were growing well — customer support and coding — making the pivot an empirical discovery rather than a strategic bet.
- Varun argues that coding agents provide approximately 7x engineering leverage at GigaML, allowing them to maintain a very small but high-quality team — and that the benefit is not just cost savings but reduced context-switching and faster shipping through single ownership of features.
- Varun contends that in the AI era, product is the dominant competitive factor over sales, pointing out that Anthropic and OpenAI do not even pay sales commissions, and that no successful AI company is winning on the strength of its sales team.
- Varun describes the core technical insight behind GigaML's agentic product: enterprise AI deployment performance fundamentally reduces to the quality of policy markdown files and the ability to iteratively improve those files to move a specific business KPI like resolution rate or CSAT.
- Varun's YC interview experience revealed that HJ (Harge) ignored the founders' prepared pitch — TAM, idea, market size — and instead evaluated them purely on their engineering credentials, admitting them conditionally on pivoting away from edtech entirely.
Topics
Full transcript available for MurmurCast members
Sign Up to Access