Giga Says Product Velocity Beat a 400-Person Rival at DoorDash

Varun Vummadi describes Giga as a company building AI agents for customer support, with customers including DoorDash, one of the biggest crypto exchanges in the US, and top telecom providers. The product is meant to change a familiar support path: a customer calls, enters an IVR or chatbot flow, and is routed to a human. In Vummadi’s telling, traditional systems deflect only about 10% to 15% of calls before a human gets involved. Giga’s AI support agents are meant to make the call itself feel “human-like” and resolve a much larger share of issues without escalation.

The target is higher. Vummadi said Giga is aiming for 90% to 95% deflection for some customers. The promise is not framed only as a lower-cost chatbot, but as a different customer experience: no hold time, faster resolution, and less routing through layers of support infrastructure.

Vummadi’s broader view of AI agents is similarly operational. In deployed enterprise systems, he says, the work “fundamentally boils down to” policies, or what he calls “the markdown file,” and the ability to keep changing that file until a business KPI moves. For support, the KPI might be resolution rate or CSAT. For other categories, he says the same fundamentals apply to compliance, IT service management, and internal support.

His example is simple: a customer might begin at 30% to 40% resolution, and the work is to improve toward 90% by iterating the agent’s policies and behavior. The model is only part of the system. The enterprise value comes from configuration, measurement, and repeated improvement against a business metric.

That view leads directly to Giga’s next product direction. Vummadi says the biggest bottleneck in enterprise AI deployments is the need for forward-deployed engineers: people who sit with customers, configure systems, make policy changes, build dashboards, attend meetings, and translate business intent into software changes. Giga is trying to build what he calls an “AI forward deployed engineer.” The system would join Slack and Google Meet, take notes, make policy changes, and help customers move metrics such as resolution rate from one target to another.

That is not separate from customer support. It is a generalization of what Giga has learned from support: enterprise AI adoption depends less on a one-time model deployment than on the speed with which a system can be configured, measured, and improved inside a real business process.

Giga did not begin as a customer support company. Vummadi and his co-founder entered Y Combinator after applying with an LLM-based edtech idea, at a moment when ChatGPT had recently launched and they were looking for something to build on top of it. Vummadi had been doing LLM research in college, including work involving transformer models such as BERT before ChatGPT’s release. He had an offer to join a leading quant firm in New York at $550,000, and he also referred to having a Stanford PhD path available to him.

Instead, he and his co-founder applied to YC. Vummadi had been reading Paul Graham essays and watching older YC Startup School material, and they decided to “give it a shot.” The timing was tight: he said they were supposed to join their jobs in three days when they finally got accepted.

The YC interview did not go the way Vummadi expected. He had prepared by speaking with former YC founders and expected questions about the idea, TAM, and related startup fundamentals. Harj Taggar, then interviewing them, did not focus on those. According to Vummadi, Taggar told them the edtech idea was not going to work and that they should pick something else.

Vummadi says he thought the interview had gone badly enough that they would not be accepted. Instead, he says YC took a bet on their engineering ability rather than on the specific idea. “Giga would not have existed without Harj taking a bet,” he said.

That early rejection of the idea became the first pivot. Taggar connected them with the Coursera COO and others who had built successful edtech companies. Vummadi says they all told him edtech was a bad idea. After about a month in the batch, the founders pivoted away from it.

The first serious replacement was not customer support. It was fine-tuning. Their own immigration complications — both founders’ B-1/B-2 visas were rejected, forcing them to participate remotely as YC returned in person — did not change the technical direction. Vummadi had read a research paper from a Databricks co-founder about caching LLMs to reduce cost. GPT-4 was expensive at the time, and Giga’s founders thought fine-tuning smaller models could be a better route than caching alone.

They open-sourced models, topped Hugging Face benchmarks, attracted inbound interest, and raised a $4 million seed round. But after about a year, Vummadi says they realized fine-tuning was “a really bad market” for their purposes. In his account, customers fine-tune primarily to reduce cost and latency, or because they need secure deployments for sectors such as insurance or healthcare. Selling into those secure enterprise markets was not mainly an engineering process, he says. It was a sales process.

The real company emerged from usage, not from a top-down market selection exercise. Among Giga’s customers, Vummadi says the two use cases growing well were customer support and coding. Giga chose support. Zepto became the first customer for that direction after Giga reached out as Zepto was scaling quickly.

Ankit Gupta pressed Vummadi on whether that meant Giga discovered the idea from customers rather than from abstract market analysis. Vummadi agreed. The support business was found inside demand that already existed.

When Giga moved into customer support, better-funded competitors already existed. Gupta named Sierra as an example of a company with famous founders and substantial capitalization in the same broad market. Vummadi’s answer was not that Giga had a sophisticated competitive thesis from the beginning. It was that they did not know much about those competitors when they started.

That ignorance was useful, or at least not fatal. Vummadi said the team’s mentality was simple: if customers were willing to pay and Giga could deliver value, they should do it. Competition did not dominate the decision.

The sharper test came after Zepto, when Giga competed for DoorDash. Vummadi says Giga had eight people and was competing against a well-funded company with roughly 400 people. Giga won.

Vummadi does not present the DoorDash deal as proof that enterprise buyers are universally indifferent to startup risk. He says Giga had an unfair advantage: YC. Garry Tan introduced him to Tony Xu, and Vummadi emphasized the trust created by both Giga and DoorDash being YC companies. The pilot still had to work. He said Giga piloted for three months, did not go down, and had good metrics. He also credited DoorDash as a meritocratic company willing to choose a small startup at unusual scale.

The DoorDash win then changed Giga’s credibility with other buyers. Vummadi says many companies now choose Giga because DoorDash and other large public-company users have already done so. The early trust advantage became referenceable proof.

Vummadi’s lesson from the deal was that there is “a lot of arbitrage” in building a great product rather than building a large sales team. Gupta framed the contrast against older enterprise software assumptions: a decade ago, many would have assumed a large enterprise would not buy software at that scale from a startup. Giga’s experience suggested that, at least for certain buyers and use cases, performance in a pilot and trust from the YC network could outweigh the startup’s small size.

Vummadi’s personal story matters because the risk was not abstract. He grew up in a small town in Andhra Pradesh. His parents were government teachers and expected him to become an engineer or a doctor. He “grinded” to get into IIT Kharagpur, where he studied electrical engineering.

The two founders were both technical, but Vummadi describes them as different types. His co-founder, he said, was third ranked in IIT and received only one B in his entire academic career. Vummadi presents him as the bookish one. Vummadi says his own grades were bad and that he spent time on Kaggle competitions because winning could make money. He says he made close to $50,000 through Kaggle and that this helped him land a high-frequency trading job. He also says he “gamed” Kaggle so much that he was banned.

Both founders turned down highly paid quant paths. Vummadi had the $550,000 New York offer. His co-founder, he said, had the highest-paying job offer in India from a quant firm. Vummadi says many people thought he was stupid to reject the offer. His father was “super mad,” and the decision caused a major fight at home. The offer would have been meaningful for a middle-class or below-middle-class family, and Vummadi says even he wondered internally whether he was doing something wrong.

His argument to his parents was that YC was a credible shot, and that if it did not work after a year or two, he could return to a job. More broadly, he frames the decision as a way to test potential rather than optimize near-term expected income. He and his co-founder wanted to see “how high we can go.” That same mindset, he says, later shaped their decision to reject acquisition offers from some of the biggest companies in the world.

Vummadi also ties this philosophy to Paul Graham’s essay “How to Make Wealth,” which he interprets as boiling down to doing a company or holding equity in something large. His co-founder, he says, is “zero motivated by money,” and that influenced the company’s willingness to keep pushing rather than sell early.

The advice is not that risk disappears. It is that the downside for capable young technical people may be less permanent than it feels. “If you have a job you will again get a job,” he said. In his view, starting and selling become more urgent when the fallback option is not the thing occupying your attention.

Vummadi is blunt about startup ideas: having them is not the hard part. “I can just go onto ChatGPT and get like 10 ideas,” he said. The hard test is whether anyone is willing to pay.

He says Giga wasted time, even after entering YC, working on “stupid ideas” that did not generate revenue or meaningful traction. For new products, Giga now tries to determine in advance whether the customer can actually pay, estimates how much the customer would pay, gets a commitment, and then builds.

The advice is especially pointed for B2B founders. Vummadi says that if a problem is important enough, people will pay with money or time. Social networks may be paid for with time rather than cash, but for B2B, if a buyer will not pay money, he treats that as evidence the team may be solving a fake problem.

That is why Vummadi puts less weight on broad market reasoning at the earliest stage. He says he does not even think founders need to care much about market size if someone is willing to pay real money for a solved problem and the value delivered is clear. Elsewhere, he acknowledges that as a company scales, differentiation becomes necessary. But for early discovery, his emphasis is on payment and value, not slides about theoretical market opportunity.

His view of geography is similarly practical. Founders should stay close to customers, wherever those customers are. For generative AI and research-based work, however, he strongly believes San Francisco is the place to be because of access to researchers and because, in his view, much of the innovation in generative AI is being driven from the Bay Area. If the customer base is primarily in India, he says, the company should be in India.

Giga’s internal operating model is meant to reflect the product thesis. Vummadi says one of the company’s values is “automate, automate, automate,” and that employees are pushed to use as much automation as possible. The company’s larger vision, as he states it, is to “automate all of the world’s work,” moving toward a generic automation builder that can automate many kinds of tasks.

The examples are routine but revealing. Vummadi says people no longer need a personal assistant for some scheduling tasks; they can use Claude to schedule meetings. Salespeople use transcripts from Gong to analyze what works or does not work against a specific competitor across many calls. His broader point is that tools such as Claude Code turn more employees into builders, letting them create workflows and extract insights that would otherwise require manual review.

Gupta asked what Giga would look like without coding agents. Vummadi estimated that the company would need six to seven times as many engineers. He emphasized that the advantage is not only cost. A smaller engineering team also reduces context switching. In his view, it is better when one person can own and build an entire thing rather than spreading the work across many people and paying the tax of context transfer.

That changes how Giga hires. The interview process explicitly tests both AI-enabled building and underlying understanding. Candidates are asked to “vibe code,” then the company removes AI access and asks them to change the code without it. The goal is to see whether they understand the code and how it works, not just whether they can prompt a tool into producing something plausible.

Vummadi admits this standard may evolve as AI models improve. If Claude can produce the whole thing, he asks, how much does the human need to understand? For now, Giga’s answer is that both abilities matter.

The company also looks for “extraordinary ability and spikiness.” Vummadi uses the founders’ own histories as examples: his unusually high-paying offer, his co-founder being third ranked in IIT and having a quant offer. Giga is looking for evidence that someone has done something in the top fraction of a percent, not necessarily a conventional résumé match.

Vummadi says he began with the wrong internal debate. He thought sales might be the most important function in the company. His co-founder disagreed. Vummadi now says he was “so wrong.”

His current bias is strongly toward builders over sellers, especially in AI. He points to successful AI companies as examples of product strength driving adoption. “Nobody uses Anthropic for the best sales team,” he said. He also claimed that Anthropic and OpenAI do not pay salespeople commissions, using that as an illustration of his view that sales is not the center of gravity in those businesses.

The claim is not that sales is irrelevant in every market. Earlier, Vummadi’s critique of fine-tuning included the realization that selling secure deployments into large insurers or healthcare companies was a sales process rather than an engineering process. But he argues that in AI, the decisive question is whether the product can deliver substantial value quickly. If it can, “everything else should follow through.”

This product-first posture also shapes how he thinks technical founders should approach business inexperience. Gupta asked whether founders without business backgrounds should worry about that deficit. Vummadi’s answer was that many buyers will purchase a product from technical founders if the buyer is the right fit. Zepto and DoorDash, in his account, did not care whether Giga had senior salespeople. They cared whether the product solved the problem.

The key is finding the right ideal customer profile: buyers who reward product performance and speed of value delivery rather than the presence of a large sales organization. That is not every buyer. But Vummadi’s experience is that those buyers exist, and a small technical team can build a company around them.