Why Two IIT Engineers Turned Down $550K Jobs To Build A Startup

Despite a disastrous YC interview that nearly ended in rejection, the founder of Giga ML credits Harj’s decision to bet on the engineers as the spark for the company, which pivoted within a month from EdTech to building AI customer-support agents used by DoorDash and a top-three telecom. Their human-like calls lift deflection rates to 60–70% versus the 10–15% of traditional AI, with top customers targeting 90–95%. The core insight is that an AI agent’s effectiveness reduces to a policy stored in a markdown file, and the real work is iterating on that file to move KPIs like resolution rate and CSAT. The founder urges founders to “burn the boats” and turn down safe offers, arguing that product and value delivered beat sales teams—as proven when an eight-person Giga ML outmaneuvered a well-funded rival to win DoorDash. Looking ahead, they are building an AI “forward-deployed engineer” to automate implementation, embedding a culture of “automate, automate, automate” where coding agents let a small team operate like one six to seven times larger.

• ▶ 0:19 The founder felt the YC interview went so poorly they expected rejection—Harsh advised them to "pick something else and work on it."
• ▶ 0:39 Giga ML builds AI agents for customer support, serving companies like DoorDash, a major crypto exchange, and a top-three telecom.
• ▶ 1:27 Giga ML's human-like conversational calls lift deflection rates to 60–70% (vs. 10–15% for traditional AI), with top customers targeting 90–95% deflection.
• ▶ 2:27 Varun did early LLM research at Stanford (pre-ChatGPT) on BERT/transformers, leading to a quant job offer and a Stanford PhD offer.
• ▶ 3:04 The ChatGPT launch catalyzed him and his co-founder to build on top of it and apply to YC on a whim—just before he was set to start his job.
• ▶ 3:55 At the YC interview, Harj dismissed their EdTech idea, ignored prepared questions, and bet on them as engineers anyway—a bet Varun credits for the company's existence.
• ▶ 6:29 After YC and edtech operators all said their original idea was bad, the founders pivoted less than a month into the YC batch.
• ▶ 8:16 They discovered organically from customers that customer support was a high-growth use case, leading them to pivot to customer service AI.
• ▶ 9:37 Despite being an eight-person team facing a well-funded competitor, they won DoorDash, proving a great product could beat a big sales team.
• ▶ 11:00 The company has scaled from early customers like Zepto and DoorDash to now working with the largest U.S. crypto exchange and many Fortune 500 companies on automating support operations.
• ▶ 11:25 The core insight: AI agents' effectiveness comes down to a policy stored in a markdown file and the ability to iterate on that markdown file to impact a business KPI—applying across almost any company or use case.
• ▶ 11:37 In practice, this means iterating on specific KPIs like support resolution rate (e.g., improving from 30-50% to 90%) and CSAT, with the same principles extending to compliance and ITSM/ITSD.
• ▶ 12:40 Turn down safe offers if you truly want to build something big; this ambition shaped their company and later led them to refuse acquisition offers from the world's largest companies.
• ▶ 13:57 Expect pushback from family when taking the startup path; he convinced his parents by showing YC's credibility and arguing he could always return to a normal job if it failed.
• ▶ 15:03 Startups fail when founders work on "stupid ideas" with no revenue; the real test is whether anyone is willing to pay real money for what you build, so get a paying commitment before building.
• ▶ 16:26 Location strategy: stay close to customers; for GenAI/research-heavy work, the Bay Area is the right hub, but if customers are primarily in India, base the team there.
• ▶ 17:12 Future product direction: building an AI "forward-deployed engineer" that joins Slack/Google Meets, takes notes, and implements changes automatically to solve enterprise AI deployment bottlenecks.
• ▶ 18:29 Internal AI culture: core value is "automate, automate, automate," with the mission to "automate all of the world's work" and employees using their own product for tasks like meeting scheduling.
• ▶ 19:13 Sales teams use Claude Code to analyze Gong transcripts, extracting insights about competitor positioning that would normally require heavy manual analysis.
• ▶ 19:46 The founder says coding agents let them operate with 6–7x fewer engineers; the real wins are no context switching, faster shipping, and avoiding momentum-killing context transfer.
• ▶ 20:20 The hiring process now requires candidates to write code, then modify it without AI access to verify true understanding; they also look for extremely rare, "spiky" top-0.1% talent.
• ▶ 21:44 You don’t need a business background to succeed—just find the right buyer and your ICP (Ideal Customer Profile), as seen with Zepto and DoorDash.
• ▶ 22:47 Product matters most, especially in AI: even Anthropic and OpenAI don’t rely on sales teams or commissions—value delivered to customers drives everything.
• ▶ 23:37 The biggest lesson is to commit fully and “burn the boats”—once you reject the safety nets, that pressure forces you to make things happen, and with AI’s low building costs, just start building.

• ▶ 0:00 Opening and Product Overview (0:00 - 1:34) - - A panicked cold open leads into what Giga ML is and how its AI customer service agents work.
• ▶ 1:34 Origin Story: Upbringing to YC Interview (1:34 - 6:08) - - Varun's upbringing, college, engineering roots, Kaggle history, ChatGPT launch, and the surprising YC interview.
• ▶ 6:08 Pivot to Customer Service AI and Winning DoorDash (6:08 - 10:58) - - From fine-tuning/EdTech to customer-led customer service AI and winning DoorDash as an eight-person team.
• ▶ 10:58 Scale, AI Agents, and Markdown-Driven KPIs (10:58 - 12:13) - - Current company scale, how AI agents use markdown policy files, and business KPI iteration.
• ▶ 12:13 Advice for Students and Startup Validation (12:13 - 16:26) - - Advice to students, parents' reaction, and validating startups through willingness to pay and charging early.
• ▶ 16:26 Location, Future Product Direction, and Internal AI Use (16:26 - 19:09) - - India/U.S. location strategy, the future AI forward-deployed focus, and internal AI automation culture.
• ▶ 19:09 AI-Powered Sales, Engineering, and Hiring (19:09 - 21:41) - - Sales using Claude Code with Gong transcripts, coding agents' impact on headcount, and the modern hiring process.
• ▶ 21:41 Founder Mindset and Parting Advice (21:41 - 24:30) - - Technical founders and business, product as the most important thing, and final advice to burn the boats and just build.