Varun Mathur left a video model that beat Gemini 2.5 Pro to chase a quieter problem: the few experts every software company leans on until they break. Unisson, his two-person YC W26 startup, wants to spread that expertise to everyone.
Inside almost every growing software company there is a small group of people who quietly hold the whole thing together. They know why the integration broke. They remember the workaround from last quarter. When a big customer stalls during onboarding, someone types their name into Slack and waits. Varun Mathur spent years watching that pattern repeat, and he came to see it as the most expensive bottleneck in modern software.
His answer is Unisson, a company he co-founded in early 2026 and now runs as CEO. It is small - two people - and it is aimed at a problem most startups skip because it is not glamorous: the grinding, hands-on work of getting customers to actually use a product. Onboarding. Migrations. Change management. Health audits. The tasks that never trend but decide whether a deal survives its first ninety days.
The ideaMathur has a phrase for the people who carry this load. He calls them the human API. Customer-facing teams - technical customer success, implementation, sales engineering - become the living interface between a product and the people trying to adopt it. Every question routes through them. Every complex setup waits on them. And because real expertise concentrates in a few heads, those few heads turn into a wall the rest of the company keeps hitting.
The consequence is predictable. Time-to-value slows. Cost-to-serve climbs. The best engineers get pulled off the roadmap to answer the same implementation questions for the tenth time. Unisson's bet is that this expertise does not have to live in three inboxes. It can be captured, learned, and handed to an agent that never sleeps and never forgets the workaround.
How it worksThe demo that gets attention is speed. Unisson's agents explore a product's interface and, in roughly fifteen minutes, learn to operate it well enough to help. A new hire needs weeks or months to reach the same point. The agents pull context from meetings, transcripts, and Slack, then execute real work end to end rather than just answering questions. They live where teams already are - Slack, text, email - not behind another dashboard nobody opens.
A built-in ticketing layer watches what customers keep asking for and where the friction lives, so the product team gets a running read on operational trends instead of a stack of one-off complaints. The pitch to a buyer is blunt: give your customer-facing team round-the-clock access to a hands-on subject-matter expert, and let them scale without cloning their three best people.
Mathur did not arrive at customer success by accident. He studied computer science and statistics at the University of Illinois Urbana-Champaign, finishing in 2019, then spent years at Ambient.ai, a fellow Y Combinator company from the Winter 2017 batch. There he led product, engineering, growth, and vision-language-model research for agent products. The headline achievement from that stretch is technical: he built a video-understanding model that outperformed Gemini 2.5 Pro on the benchmark.
He also learned to sell. At Ambient.ai his AI reached Fortune 5 companies, which is a different kind of education - less about model architecture, more about what enterprises actually need before they trust software with real work. Both lessons show up in Unisson. The model research explains why he believes an agent can learn a product fast. The enterprise selling explains why he chose a problem the Fortune 5 feels in its bones.
Unisson is not a solo run. Mathur's co-founder and CTO is Tom Achache, who led the Perception team at Chef Robotics, where he helped deploy more than fifty production robots that served over eighty million meals. Before that, the two overlapped at Ambient.ai. Starting a company with someone you have already shipped hard systems with removes a variable most first-time founders never resolve: whether the other person actually delivers under pressure.
The pairing is telling. One founder spent years teaching machines to understand the messy visual world; the other spent years getting robots to act reliably in real kitchens. Unisson sits at the intersection - software that has to both understand an unfamiliar product and take dependable action inside it. Their YC partner for the Winter 2026 batch is Diana Hu.
The betThere is a version of Mathur's career where he keeps optimizing video models and chasing benchmark records. He walked away from that toward something harder to demo on stage: migrations and change management. It is a deliberate choice. Software got dramatically easier to build over the last decade. It did not get easier to adopt. That gap - between shipping a product and getting a customer to truly use it - is where Unisson has planted itself.
The company is early. It launched publicly through Y Combinator in early 2026 with a small seed round and a two-person team. What it has instead of scale is a sharp thesis and two builders who have seen the bottleneck from the inside. If they are right that rare expertise can be turned into a product that answers at two in the morning, the customer-success org of the next decade looks very different from today's.
For now, Mathur is doing what early founders do: learning fast, shipping faster, and betting that the least glamorous corner of B2B software is exactly where the leverage hides.