Ai Daniil Bekirov likes doors that look slightly closed. University admissions, startup internships, Y Combinator, a seat beside Sam Altman - each became less a verdict than an invitation to ask a better question. His current question is aimed at the most guarded door in modern office life: who is allowed to change the software?
Most companies have a familiar ritual. Someone in marketing wants a line of copy changed. Operations needs an internal form adjusted. A designer can see the spacing problem from across the room. The request joins a queue, acquires a ticket number, and waits for an engineer who knows the repository, the local setup, the permissions, and the obscure commands that make the whole machine hum. The change may take five minutes. Reaching those five minutes can take a week.
Bekirov founded Sparkles to shorten that approach without pretending software has become consequence-free. A teammate describes a change in ordinary language. A coding agent works against the company's existing repository inside an isolated cloud sandbox. Commands and edits stream live. The result comes back through GitHub as a pull request, where reviews, branch protections, and automated checks still apply. The non-engineer gets agency. The engineer keeps the veto. Production is not used as a rehearsal room.
A first name fit for the moment
There is an unavoidable joke in a 21-year-old named Ai building an AI company. Bekirov gets there first. His biography has long opened with “Legally AI,” though he usually introduces himself as Dan. When an interviewer announced that he had interviewed “a guy named AI” out of a window at Y Combinator, Bekirov replied with the gentlest possible correction: “I'm Dan btw.” His GitHub profile offers another compact manifesto: “I don't like Java.”
The punch lines disguise a serious appetite for making things. Born on May 28, 2005, and from Crimea, Bekirov says he built a GPU mining rig at 13. By 2025 he was a first-year computer science student at University College London, struggling to land an internship and newly experimenting with Cursor. A hackathon gave him a faster form of feedback. He won one in February. Then he won nine more.
One of those trips took him to Paris for OpenAI's first EMEA Viva Le Hack. With Joséphine Parquet and David Gelberg, he built Nerd Radar, a talent search tool that connected arXiv papers with Google Scholar profiles so recruiters could find researchers who did not cultivate LinkedIn pages. The team placed third and received $10,000 in OpenAI credits. The project was a small preview of Bekirov's recurring instinct: find capable people who are invisible to the usual interface, then make their work legible.
Hackathons led to startup work. At Iterate, founded by former Monzo chief technology officer Jonas Templestein, Bekirov sharpened his TypeScript skills. Structured AI then hired him as its first employee. Co-founder Raymond Zhao wrote that the freshman shipped more than ten features in his first week, including a local-development tool, an agent system designed to use fewer tokens, and an evaluation set. Zhao's explanation was one word: agency.
The application, then the edits
Bekirov's route into Y Combinator was gloriously untidy. He applied on the final day with what he later described as an idea and a dream. He pivoted one day before the interview. He got into the Winter 2026 batch, then pivoted again into Sparkles. The polished founder myth prefers a straight line. His version has version history.
Leaving UCL was less impulsive than the announcement made it sound. After working inside fast startups, he returned to lectures and found the pace newly hard to accept. Hackathons had shown him that he could build; startup jobs had shown him that output could outrun seniority. At 20, he moved from London to San Francisco and wrote about working at 2 a.m. with three Claude Code sessions running on a second monitor. His relatives wanted him to finish the degree. He wanted to test the unusual opening in front of him.
Then a cold LinkedIn message from OpenAI produced a scene almost too neatly scripted. At a January 2026 town hall, Bekirov asked where Sam Altman would sit and took the adjacent chair. When the microphone arrived, he asked two questions. Should he return to university? And did Altman angel invest? Altman told him that an ambitious AI builder could make his own choice and return later if needed. On the investment question, the answer was no. Bekirov sent the exchange to his parents. Even a rejected pitch can carry useful cargo.
A city measured in lucky collisions
Bekirov had rehearsed the move in an essay months earlier. During a 2025 visit to San Francisco, he complained about the wind, $9 coffee, and a culture that could assign too much value to a glittering résumé. He also became fascinated by the city's density of accidental meetings. After an exhausting climb to a taco shop, he and his friends struck up a conversation with four MIT students headed to the same place. Bekirov turned the encounter into mock mathematics: San Francisco's serendipity coefficient exceeded London's.
The observation was playful, but the conclusion stuck. Put more golden marbles in the bag, he argued, and you improve the chance of drawing one. He valued people who stayed interested before hearing the impressive part of a biography. Yet he also wanted to live near the builders, investors, and friends who made ambitious action feel ordinary. By 2026, Sparkles was operating in San Francisco, and his personal site listed angel investments in Touchmark, Byteport, Forum, Flick, and Zephyr Fusion.
Hackathons gave Bekirov a place where proof could arrive before a long résumé.
Internships, YC, and a famous town hall question all began before certainty did.
Sparkles expands who can propose software changes without erasing engineering judgment.
His San Francisco thesis is environmental: useful collisions become likelier when builders cluster.
The product is permission, carefully packaged
Sparkles began with a memorable comparison: Lovable for existing projects. The distinction matters. Blank-page app builders create something new. Sparkles enters repositories that already have customers, conventions, secrets, and scars. The service connects to GitHub, prepares the environment, and allows colleagues to run Claude Code, Codex, or OpenCode through a shared interface. Secrets can sync from dedicated vaults. Every session remains visible. A proposed change returns to the workflow engineers already trust.
The company's own language has widened from “make everyone on your team an engineer” to “agents for your whole team.” That is a quieter and more accurate ambition. A marketer does not need to become an engineer to fix a price on a page. An operator does not need a new identity to adjust an internal tool. The point is to make useful intent executable, then make the execution inspectable.
Illustrative friction from idea to review
That framing also explains why the pull request is more interesting than the prompt box. AI can generate code in many places. Companies still need a social mechanism for deciding whether generated work belongs in the product. The pull request carries context, responsibility, checks, and a place to say no. Sparkles is betting that the old review object can become the bridge to a much broader group of contributors.
Bekirov is equally candid about distribution. In an interview, he argued that early startups should focus on product and use organic posts to begin conversations before spending on advertising. A joke about a San Francisco apartment building's fundraising-themed Wi-Fi name drew roughly 10,000 views overnight and placed the Sparkles logo beside every impression. More recently, he announced a mock pivot into donut delivery while linking to a real donut site, then reported demoing Sparkles at Cloudflare. His feed treats company building as a serialized comedy with working software hidden inside.
The next compile
On his 21st birthday, Bekirov compared the year with the one before it. At 20, he had been delighted to secure a startup internship. At 21, he had sent an offer to Sparkles' first hire at more than twice his former salary. He also wrote that Sparkles had raised $3.5 million, with backing listed from Y Combinator, Moonfire, scout funds, and people connected to OpenAI, Anthropic, a16z, General Catalyst, Cloudflare, and other firms.
The money raises the stakes, but it does not settle the product question. Teams will decide whether they trust coding agents with real repositories. Engineers will decide whether incoming pull requests save review time or manufacture more of it. Non-engineers will decide whether watching an agent work feels like power or merely a nicer waiting room. Sparkles has to earn all three constituencies at once.
Bekirov's advantage may be that he is not precious about the first draft. His brief career is a string of fast experiments linked by a durable preference for action: build, show, ask, revise. He once summarized the playbook as “Build. Iterate. Show. Build.” The phrase could describe a hackathon, a YC application, or the Sparkles interface itself.
There is wit in choosing the pull request as the instrument of workplace liberation. It is bureaucracy with syntax highlighting, a proposal designed to invite objections. But that modesty may be precisely the point. Sparkles is not asking a company to abandon its controls. It is asking whether those controls can admit more useful proposals.
For a founder who built his life by making asks before he possessed conventional permission, the product is unusually personal. The door stays on its hinges. The lock still works. More people get a key to the lobby - and a chance to show what they would change.