For six months, David Gu’s nights followed two possible scripts. In one, the recording infrastructure failed and a pager pulled him awake. In the other, nothing failed, but he dreamed that the pager had gone off anyway. The distinction was not especially restful. He had written the first version of the call-recording software, and he was the person responsible when it buckled under real meetings.
The recording product he and Amanda Zhu were building did have users. Thousands of them. It processed millions of hours of video. Its intended value lived in the polished layer above the machinery: analysis, insight and search for user researchers and product managers. Yet about 80 percent of the engineering effort went into capturing the calls reliably. If that work succeeded, customers saw nothing. If it failed, they lost the meeting they had expected to keep.
Infrastructure gave the founders a cruel bargain: do it perfectly and receive no credit, or do it imperfectly and make everyone angry. The burden followed Gu for three years, repeatedly pulling him away from feature work to stabilize a system that had to wake up for the morning rush, ingest live media and preserve data that could not be recreated.
“When you’re a product company, you can’t win with infrastructure.”David Gu
Eventually, he and Zhu looked at the burden differently. Other software companies wanted to build products on top of meetings, too. They would all need the same awkward integrations and real-time reliability. The problem consuming Recall’s predecessor might be more valuable than the application it was supporting.
The useful part of being number five
Gu and Zhu met the startup world early. Gu arrived at the University of Waterloo in 2018 to study software engineering, carrying a Schulich Leader Scholarship and a history of building things. Before university, he had started LearnVR, formed two school clubs, made small educational programs and joined hackathons. A virtual-reality project helped his team win at MIT’s Reality Virtually Hackathon. Two years later, at 19, he left Waterloo during his second year and entered Y Combinator’s Winter 2020 batch.
The first company was called Perfect Recall. During Y Combinator, the founders had conducted so many user interviews that recordings accumulated on their hard drives. They built a tool for capturing and revisiting those calls. The need was real, but the category filled quickly. After three years, their product had customers who cared about it and a position Gu described bluntly: number five in a crowded market.
That was uncomfortable because it was neither a clean failure nor an obvious success. People relied on what they had made. Shutting it down meant disappointing them. Continuing meant accepting a future as another option in a busy field. The founders decided that building something new and useful mattered more than preserving the identity of the original product.
They handled the transition with a revealing bit of engineering pragmatism. The old application stayed online for six months so its customers could migrate. Meanwhile, the founders extracted its backend into Recall.ai, then refactored the old product to run on the new API. Their first customer was the company they were leaving behind.
A Black Friday at the top of every hour
Meeting infrastructure has a shape. People do not distribute their conversations evenly across a day. They click “join” near the hour and half-hour, with a pronounced surge around the beginning of the workday. Gu compares the traffic pattern to Black Friday, except the event repeats daily. Capacity must appear quickly enough to join calls, handle raw audio and video in real time, and retain the precise meeting that each customer requested.
The appeal of an API is abstraction. A developer sends a meeting URL; the infrastructure handles platform differences, bot behavior, media streams, transcripts and metadata. Recall.ai later widened the entry points. Its Desktop Recording SDK captures a meeting without a bot appearing in the participant list. The company has also moved toward phone and in-person conversations. The product line follows the thesis: important business context is spoken across many surfaces, while AI systems are only as useful as the context they can access.
In September 2025, Recall.ai announced a $38 million Series B led by Bessemer Venture Partners at a reported $250 million valuation. By March 2026, Bessemer described the platform as serving more than 3,000 companies, including HubSpot, ClickUp, monday.com and PagerDuty. Those names are useful because they are not merely “meeting recorder” companies. They suggest conversation data becoming an ingredient inside broader software.
“You need to understand the entire chain of causality from your infrastructure to the end user value.”David Gu
The engineer learns to ask
The technical pivot required a personal one. Gu and Zhu had no sales background. They learned by handling hundreds of calls themselves and closed the first $2 million in revenue before hiring a salesperson. Gu’s lesson from those early conversations is almost comically direct: ask for the deal. Name the commitment. Put a date on it. Waiting politely for momentum does not create momentum.
Recall.ai also ignored the usual script for developer products. For roughly three years, customers had to speak with someone before buying. Self-serve would have reduced friction, but it also would have removed the founders from the conversation. The calls exposed what people were building, why their users wanted it, which constraints mattered and how the infrastructure influenced the final experience.
Keep builders close to buyers long enough for architecture and demand to correct each other. Revenue records the sale. Conversation explains why it happened.
That approach fits Gu’s writing. On his small personal site, startup questions become clean systems. Product-market fit has three conditions: an important enough problem, the number-one solution for a defined customer, and a repeatable way to find more customers. Startup value becomes probability multiplied by payoff, plus progress. His “Builder’s Fallacy” argues that writing the initial code can be a fraction of the real job; adoption, measurement, feedback, documentation and support swallow the rest.
The essays are concise, but they carry scars from the company’s first years. Recall.ai began when the founders noticed that code was not the full product. Reliability, customer migration, pricing, selling, documentation and support all sat between a working system and a useful one. Gu’s engineering instinct did not disappear when he became CEO. It expanded the boundaries of the system.
Five years before it felt like it worked
Gu enters Waterloo to study software engineering.
At 19, he leaves during his second year and joins Y Combinator W20.
The founders launch Recall.ai and make their old product its first customer.
A $10 million Series A arrives as the company supports more than 300 customers.
The Desktop Recording SDK launches, followed by a $38 million Series B.
Recall.ai is reported to serve more than 3,000 companies.
Gu has said the company only began to feel as if it was working in mid-2024, nearly five years after the founders started. Before then, the operation was intentionally frugal. The two founders handled engineering, product, sales and marketing themselves. For years, external validation was sparse and the work was repetitive. The lesson he offers founders is “don’t give up,” but he adds the appropriate caveat: survivors are the people available to give that advice.
His more interesting observation concerns failure. Before a startup, a strong student or employee receives frequent signals of competence. Inside an early company, those signals vanish. Most sales calls end without a sale. Most experiments are inconclusive or wrong. Gu says he now seeks some of that feeling because it is where learning happens. Sports helped him understand the bargain: a game with no possibility of losing would not be worth playing.
The pager story is compelling because it ends with neither conquest nor escape. The difficult system did not become easy. Its demands became concentrated inside a company willing to own them for everyone else. Recall.ai sells the quiet outcome Gu once struggled to maintain: a developer builds the visible product while the recording simply arrives.
There is a practical idea to steal from that journey. Inspect the work your team resents doing over and over. Ask whether it is unique to you, or whether an entire market is reluctantly rebuilding the same thing. The painful 80 percent may be overhead. It may also be a product waiting for a boundary, a price and someone willing to keep the pager.