Key Moments
Why Two IIT Engineers Turned Down $550K Jobs To Build A Startup
Y CombinatorWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
AI agents can achieve 60-70% customer support deflection rates, but reaching 90%+ requires overcoming the 'forward deployed engineer' bottleneck, demanding AI adapt to complex business KPIs.
Key Insights
GigaML's AI agents can achieve 60-70% deflection rates in customer support, aiming for 90-95% for top customers, a significant improvement over traditional 10-15% rates.
Varun turned down a $550,000 job offer from a top quant firm to start GigaML, influenced by Y Combinator's ethos and a belief in their engineering capabilities.
The company pivoted multiple times, initially exploring edtech before focusing on fine-tuning LLMs, which led to topping Hugging Face benchmarks and raising a $4 million seed round.
GigaML won a major contract with DoorDash against a 400-person company with their eight-person team, highlighting the arbitrage in building a superior product over a large sales team.
The key bottleneck for widespread enterprise AI adoption, according to GigaML, is the need for 'forward deployed engineers,' which they are developing an AI solution to address.
The founders believe that for AI companies, product excellence is paramount, overriding the need for extensive sales teams, as demonstrated by the success of companies like Anthropic and OpenAI.
Revolutionizing customer support with AI agents
GigaML is building AI agents that fundamentally transform customer support. Unlike traditional systems that rely on IVR or chatbots with a mere 10-15% deflection rate, GigaML's AI agents provide human-like interactions, achieving deflection rates of 60-70%. Their ambition is to push this figure to 90-95% for their top-tier customers. This technology aims to eliminate customer hold times and ensure rapid issue resolution, offering a significantly improved experience. The company serves major clients including DoorDash, a leading US crypto exchange, and a top-three global telecom provider, showcasing the scalability and effectiveness of their AI solutions.
The allure of Y Combinator and a rejected $550K offer
Varun Vummadi, co-founder of GigaML, revealed that he and his co-founder turned down lucrative job offers, including a $550,000 package from a top New York quant firm, to pursue their startup dreams. This decision was heavily influenced by their previous experiences and reading about Y Combinator. They were particularly inspired by Paul Graham's essays and the YC application process. Despite having been accepted into YC for an edtech idea, the partners, particularly H.J., pushed them to pivot to a problem with greater potential, emphasizing their engineering talent rather than a specific business idea. This guidance was pivotal, as Varun admitted to being 'panicked' and unprepared for typical startup interviews but trusted YC's belief in their engineering capabilities.
From research to a 2014 YC startup school call
Varun's journey into technology began in a small town in Andhra Pradesh, India, where his parents were government teachers. Driven by the aspiration to become an engineer, he gained admission to IIT Kharagpur for electrical engineering. His initial years were more social due to the COVID-19 pandemic. However, in his third year, he became deeply involved in LLM research at Stanford. This research on transformer models like BERT, predating the widespread impact of ChatGPT, ignited his passion and foresight. This academic rigor and early exploration in AI laid the groundwork for his entrepreneurial ambitions long before GigaML took shape.
Pivoting to LLM fine-tuning and early traction
Following H.J.'s advice to pivot from edtech, Varun and his co-founder explored new avenues. They had prior experience in fine-tuning LLMs, a skill that became their next focus. At the time, GPT-4 was prohibitively expensive, and they investigated methods to reduce costs, including fine-tuning smaller models. They open-sourced several fine-tuned models, which gained significant traction and topped Hugging Face benchmarks. This success attracted considerable attention, leading to numerous inquiries and ultimately a $4 million seed funding round, demonstrating the market's interest in efficient LLM deployment.
Discovering product-market fit through customer needs
The fine-tuning venture, while successful in raising capital, proved to be a challenging market with limited core use cases focused on cost reduction and speed. Varun and his co-founder identified that the most robust growth areas for their LLM solutions were in customer support and coding. This realization stemmed directly from observing their clients' needs and engagement. Zept, their first customer, was scaling rapidly and utilized their fine-tuning capabilities, validating this market direction. This organic discovery, driven by customer traction rather than preconceived notions, guided their strategic pivot towards building AI agents specifically for customer service.
Winning against giants: The DoorDash case study
A defining moment for GigaML was securing DoorDash as a client, despite competing against a 400-person, well-funded company. GigaML, at the time, had only eight employees. This victory underscored their core philosophy: build a superior product, and you can outcompete larger entities focused heavily on sales. Varun noted that they were initially unaware of competitors like Cera and Tech1. Their approach to the Doordash deal was straightforward: if customers are willing to pay and they can deliver significant value, competition becomes secondary. This success validated their strategy of prioritizing product excellence and demonstrated that large enterprises are increasingly willing to trust smaller, agile startups with critical functions.
The 'AI forward deployed engineer' bottleneck and solution
Varun identified a significant bottleneck in enterprise AI adoption: the reliance on 'forward deployed engineers' who must work closely with clients to configure and manage AI solutions. To address this, GigaML is developing an AI-powered forward deployed engineer. This AI will integrate with communication tools like Slack and Google Meet, automatically capture requirements, and implement policy changes or configure dashboards. This innovation aims to dramatically simplify and accelerate AI deployment for businesses by automating the complex configuration and iteration process currently handled by human engineers, thereby driving broader AI adoption.
Building the future: Automate, automate, automate
GigaML's internal culture and operations are heavily shaped by their core value: 'automate, automate, automate.' Their mission is to automate the world's work, and they practice this by integrating automation tools across all functions. For instance, sales teams use AI to analyze competitor strategies from transcripts, and engineers leverage coding agents to reduce their numbers by an estimated six to seven times. Varun believes that AI tools enable a smaller, more potent team to achieve greater output, reducing context switching and enhancing ownership. Their hiring process reflects this, seeking individuals with 'extraordinary ability and spikiness,' often demonstrated by exceptional achievements or unique skill sets, ensuring a team capable of extreme productivity and innovation.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●People Referenced
GigaML Startup Advice Cheat Sheet
Practical takeaways from this episode
Do This
Avoid This
Common Questions
GigaML builds AI agents for customer support, aiming to provide a more human-like and efficient experience than traditional IVR or chatbot systems. Their goal is to significantly increase deflection rates and reduce customer wait times.
Topics
Mentioned in this video
A major food delivery company that is a client of GigaML, highlighting GigaML's ability to handle large enterprise contracts.
An AI company mentioned as an example of success driven by product rather than sales.
An AI company mentioned as an example of a successful AI product where sales is not the primary focus.
A company that builds AI agents for customer support, working with large companies like DoorDash and major crypto exchanges.
GigaML's first customer for their customer support AI, which was scaling rapidly.
Mentioned as the source of a research paper on LLM caching that inspired the founders' fine-tuning approach.
A competitor in the AI customer service space that existed when GigaML decided to focus on this area.
Mentioned as a competitor that GigaML was unaware of when they signed Zepto.
A transformer model mentioned in the context of LLM research during the founder's college years.
A startup accelerator that accepted the founders after they pivoted from their initial edtech idea.
A platform where GigaML open-sourced models and topped benchmarks, gaining significant traction.
A video conferencing tool used by GigaML's AI forward deployed engineer to collect data for improvements.
A communication platform used by GigaML's AI forward deployed engineer to gather information and make changes.
Met with the COO of Coursera to discuss the edtech idea, who along with others, advised against it.
An AI model that launched around December, exciting the founders and inspiring them to build on top of it.
More from Y Combinator
View all 593 summaries
68 minInference, Diffusion, World Models, and More | YC Paper Club
47 minInside YC's AI Playbook
51 minWhy Good Companies Go Bad (And How to Stop It)
29 minWhy Zepto's Aadit Palicha Turned Down Stanford to Deliver Groceries
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free