🚀 How an 8-Person AI Startup Beat a 400-Person Competitor for DoorDash

💼 The $550K Job Offer That Never Got Cashed

Not many 21-year-old engineers turn down a $550,000 job offer from a leading New York quant firm to build a company. Fewer still do it after being admitted to a Stanford PhD program. Yet Varun, co-founder of GigaML, did exactly that — and his parents were furious.

"My dad was super mad to be honest," Varun recalls. "It was a big fight thing at home." But the decision was driven by something deeper than financial security: a philosophy of taking shots and seeing how high you can climb.

Today, GigaML builds AI agents for customer support, working with some of the world's largest companies including DoorDash, a top-three global crypto exchange, and major telecom providers. The company has transformed customer support deflection rates from the traditional 10-15% to 60-70% — with a roadmap targeting 90-95% for top customers.

"It's just a better experience — you never need to be on hold again. You can just call and get the issue resolved really fast."

🎓 The IIT Hacker Who Gamed Kaggle

Varun's journey began in a small town in Andhra Pradesh, where both his parents were government teachers. Like many Indian students, he grinded his way into IIT (Indian Institute of Technology), landing in electrical engineering. But his path diverged from the typical book-smart trajectory.

While his co-founder was third-ranked in the entire IIT and only received one B in his entire academic career (due to a 10-mark penalty for signing attendance for a friend), Varun took a different approach. "My grades were bad. I never really studied that good," he admits.

Instead, Varun became obsessed with Kaggle competitions — primarily because you could make money by winning. He earned approximately $50,000 through competitions and eventually gamed the system so effectively that "they banned me." This unconventional path landed him high-frequency trading jobs and, more importantly, deep technical expertise in machine learning.

During his third year, Varun began research in LLMs at Stanford, working on transformer models like BERT in the pre-ChatGPT era. When ChatGPT launched in December, everything changed.

🚪 The YC Interview That Went "Horribly Wrong"

Varun and his co-founder applied to Y Combinator with an edtech idea built on LLMs. They prepared extensively, talking to numerous YC founders about TAM (total addressable market), pitching strategies, and standard interview questions.

Then they met Harge Taggar (a YC partner).

"I was just panicked because I never prepared for anything," Varun recalls. "I prepped so much... They all told me 'what's your idea, what's your TAM,' but Harge didn't ask about any of those things."

Instead, Harge told them directly: "This is edtech, it's not going to work. Pick something else." He arranged calls with the Coursera COO and founders of successful edtech companies. They all confirmed it was a bad idea.

Varun thought the interview went "so horrible that we're not going to get in." But Harge saw something else: "You guys are really good engineers. Just pick something else and work on it."

They got in. And they had three days before they were supposed to start their jobs.

"You would not have existed without Harge taking a bet."

🔄 The Pivot That Almost Never Happened

After a month in the YC batch, Varun and his co-founder pivoted. Their visa situation was complicated — both their B1 visas got rejected, and this was the first time YC was transitioning back to in-person after going fully remote.

They landed on LLM fine-tuning. At the time, GPT-4 was extremely expensive, and they'd read a research paper from one of the Databricks co-founders about caching LLMs to reduce costs. They thought fine-tuning smaller models might work even better.

The approach worked. They open-sourced models, topped Hugging Face benchmarks, gained traction, and raised a $4 million seed round.

But there was a problem: fine-tuning is a terrible market.

"The only reason you want to fine-tune is to reduce cost and make it faster," Varun explains. The only other use case — highly secure applications in insurance or healthcare — required complex sales processes, not engineering solutions. "It's a sales process, not an engineering process."

After about a year, they noticed something: the only two use cases among their customers that were growing well were customer support and coding. They chose customer support.

Their first customer? Zepto, the Indian quick-commerce startup that was scaling aggressively.

🥊 Eight Engineers vs. 400: The DoorDash Battle

When GigaML decided to focus on customer support, they didn't realize companies like Sierra (founded by well-known, well-capitalized founders) already existed in the space.

"We didn't know Sierra and Tech1 existed when we signed Zepto," Varun admits. "We didn't think of competition much."

But the real test came with DoorDash.

GigaML — then a team of just eight people — went head-to-head with a competitor that had 400% more funding and hundreds of employees. They won.

"As you know, DoorDash is one of the massive support logos," Varun notes. "That's when we truly realized there's a lot of arbitrage in actually building a great product rather than a sales team."

How did a tiny startup convince a public company to trust them?

1. YC Network Effects: YC partner Gary Tan introduced Varun to DoorDash CEO Tony Xu. Both companies were YC alumni, creating inherent trust.
2. Technical Excellence: They piloted for three months without going down, and all metrics were strong.
3. Meritocracy: "DoorDash is a very meritocratic company," Varun emphasizes. "I really kudos to them. It's a company at that scale picking a small company would be hard."

Since then, DoorDash's endorsement has opened doors. "Now a lot of companies pick us because of DoorDash and a lot of other big public companies."

📐 The Markdown File That Changes Everything

After working with Fortune 500 companies and major crypto exchanges, Varun has identified the core mechanics of AI agents:

It all boils down to two things:

1. Policies (the markdown file that defines behavior)
2. Iteration (how you adjust that markdown to move business KPIs)

For customer support, the key metrics are resolution rate and CSAT (customer satisfaction). "How can you go from 30-40% resolution rate to 90%? How can you iteratively improve the markdown to get there?"

This same fundamental structure applies across compliance, IT service management, and internal support — areas where GigaML is now expanding with Fortune 500 pilots.

"It's fundamentally boiled down to markdown and how can you iteratively improve markdown to move a KPI."

🤖 The Forward-Deployed Engineer Problem

The biggest bottleneck in enterprise AI deployment isn't the technology — it's the forward-deployed engineer. Every enterprise deployment requires engineers to sit with customers, configure systems, make policy changes, and spin up dashboards.

GigaML's next evolution? Building an AI forward-deployed engineer.

"We're trying to build an AI forward-deployed engineer," Varun reveals. "Our AI FDE is going to join on Slack, join Google Meets, take all the notes, and do the changes automatically."

The vision is clear: move resolution rates from 40% to 60% to 90% without human intervention in the iteration loop.

⚡ Automate, Automate, Automate

One of GigaML's core values is "automate, automate, automate." The company's broader mission? Automate all of the world's work.

Internally, GigaML practices what it preaches:

• Scheduling: No personal assistants needed — Claude runs on OpenAI and schedules meetings automatically
• Sales Intelligence: Sales teams pull transcripts from Gong and use AI to analyze what works against specific competitors across multiple deals
• Engineering Productivity: Without coding agents, Varun estimates GigaML would need six to seven times more engineers

"That's the one thing I love about Claude and code. It turned a lot of people into builders," Varun says. "People are using it innovatively to drive specific insights that would have taken humans going through a lot of things."

It's not just about cost — it's about context and speed. "It's better for you to own the thing and build the entire thing rather than having a lot of people working on it. You can just ship much faster, and context transfer actually kills a lot of things and slows down things."

🎯 Hiring for the AI-Native Era

GigaML's interview process reflects this AI-first approach. Candidates are asked to:

1. Vibe code (write code with AI assistance)
2. Then have AI access removed
3. Modify the code without AI

The goal? Ensure engineers understand how code works, not just how to prompt AI systems.

Beyond technical skills, GigaML looks for extraordinary ability and spikiness — being in the top 0.1% at something. For Varun, it was one of the highest job offers in his class and Kaggle success. For his co-founder, it was being third in the entire IIT and receiving the highest-paying job offer in India from a quant firm.

💡 Lessons for Builders: The Anti-Business Plan

Varun's advice for young founders challenges conventional wisdom:

1. Ignore Your Idea (Find Your Customer)
"It's never about the idea. It's about if somebody is willing to pay you money for it."

GigaML spent too long working on ideas that generated no revenue. The breakthrough came when they focused exclusively on whether customers would pay real money for the value delivered.

"I don't even think you need to care about market to be honest. Is somebody willing to pay real money if you solve the problem for the value that you delivered."

2. Builders > Sellers (Especially in AI)
"This is the mistake I made while starting the company. I thought sales was the most important thing. I was so wrong."

Looking at successful AI companies: "Nobody uses Anthropic for the best sales team. Anthropic and OpenAI don't even pay sales people commissions. With AI, product is the most important thing."

3. Find the Right Buyer
"There are a lot of people who will buy your product without a business background. You just got to find the right buyer."

Zepto and DoorDash didn't care about sales teams or business pedigree. They cared about whether the product worked.

4. Burn the Boats
"Things get really real if you burn the boats. When the company was not working, me and my co-founder were thinking, 'Oh my god, we rejected all these job offers. What are we going to do?' It actually forces you to make things."

5. Just Start Building
"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 money."

🌍 The SF vs. India Question

As part of a new generation of Indian-origin founders splitting time between San Francisco and Bangalore, Varun has a clear framework:

"You should just stay close to customers wherever you are. But if you're doing anything closer to gen AI and very research-based things, I strongly think SF is the place — the amount of access you get to researchers is insane. Almost all the innovation in the gen AI field is getting driven in the Bay Area alone compared to India."

But if your customers are primarily in India, that's where you should be.

✨ The Philosophy Behind the Risk

Looking back, Varun's decision to reject the $550,000 offer and the Stanford PhD wasn't just about startups. It was about potential.

"For me and my co-founder, we just wanted to give it a shot. We thought, 'Let's just see how high we can go.'"

His co-founder is "fundamentally a zero-motivated-by-money person." The drive came from something else: reaching their potential and pushing to see what they were truly capable of building.

Varun also credits Paul Graham's essay on wealth, which argues that the path to significant wealth is through equity in something big, not salary.

When you come from a middle-class family, the expectations are different. The safe path is appealing. But Varun's philosophy has remained consistent: "Just take a shot and see if it works out or not."

And with AI, the cost of taking that shot has never been lower.

"Even if I don't succeed in a year or two, I can just go back to the job and do the thing again. If you have a job, you can get a job. It's not going to go anywhere."

The bottom line: GigaML's story isn't about having the perfect idea or business background. It's about recognizing an inflection point (ChatGPT's launch), finding customers willing to pay (Zepto, DoorDash), building a product that actually works (60-70% deflection rates), and constantly iterating on a simple but powerful framework (markdown policies → business KPIs).

In an era where seven times fewer engineers can build the same product thanks to coding agents, and where eight people can beat 400 with superior product execution, the arbitrage opportunity isn't in sales or pedigree.

It's in building things that work — and finding the customers who will pay for them.