He worked on navigation for robots in the dark of the Moon. Now he is building AI agents that find their way through the dark of the B2B marketing funnel.
Bhairav Mehta runs CharacterQuilt, a San Francisco company with a quietly subversive idea: most of marketing is not creativity. It is operations. Audience configuration, automation flows, list hygiene, the seventeenth round of email edits. Roughly 80% of the time it takes to launch a campaign is the boring part. CharacterQuilt builds computer-use agents to do exactly that boring part - the founder's words for it are blunt: a brain that learns your brand, and agents that operate the tools you already pay for.
The pitch lands because it is specific. Mehta likes to say that a campaign that once required three agencies, ten tools, and six weeks now gets built, designed, and deployed inside an hour. You submit a brief. You get back a campaign that already ran. The unglamorous middle - the part nobody puts on a portfolio - simply disappears.
He did not arrive at this by theory. He and co-founder Clint Burgess, who ran growth at Bloomreach for a decade, started CharacterQuilt as a full-stack marketing agency and did the work by hand first. Only after they understood every tedious step did they start handing those steps to software. It is a researcher's move dressed up as a business plan: run the experiment manually, measure the pain, then automate the variable that hurts most.
Role: Co-Founder & CEO, CharacterQuilt
Based: San Francisco, California
Before this: Founder/CEO of Buzzle AI (YC S21)
Research past: NASA JPL, NVIDIA, Mila, MIT
Batch: Y Combinator, Spring 2026
Interview real buyers. Distill them into AI personas that predict pain points and priorities. Point agents at your existing stack. Let the campaign know the customer before the customer knows you.
Marketing campaigns that used to take 3 agencies, 10 tools, and 6 weeks are now being built, designed, and deployed in an hour.
The clever part is not the automation. It is what the automation is pointed at. CharacterQuilt starts with the customer's actual voice, then lets that voice steer everything downstream.
Talk to real buyers. Capture first-party, qualitative voice-of-customer data - the stuff surveys flatten.
Distill those conversations into AI personas that predict pain points, priorities, and objections.
Agents draft messaging, audiences, and automation flows tuned to what those personas actually care about.
Computer-use agents operate the existing stack and ship the campaign - no agency, no six-week wait.
Same brief. Same channels. The difference is who does the operations - a team of people, or a fleet of agents.
Speed is not a vanity metric here. When a campaign costs six weeks, marketers ration their ideas. They run a handful of safe bets and pray. When a campaign costs an hour, the economics flip: you can run dozens of micro-campaigns, each one personalized to a specific persona, and let the data tell you which one was right.
That is the unspoken ambition behind CharacterQuilt - not faster versions of today's campaigns, but more of them, smaller, sharper, and aimed at people the software already understands. Mehta is building for a world where testing your marketing costs almost nothing.
He crushed it in the lab, then took that genius and applied it to the real world.
Before marketing, Mehta spent his twenties chasing problems that were hard for the sport of it. He studied computer science and applied math at the University of Michigan, then earned a master's at Mila, the Montreal lab that helped make deep learning respectable, advised by Liam Paull and Christopher Pal. He interned in the Computer Vision Group at NASA's Jet Propulsion Laboratory and in NVIDIA's Seattle robotics lab. His research touched a recurring theme: how machines learn to handle a world they have never seen before.
His papers have the flavor of someone who enjoys the deep end. Active Domain Randomization (CoRL 2019) asked how to train robots across deliberately scrambled simulated worlds so they survive the real one. A User's Guide to Calibrating Robotic Simulators (CoRL 2020) is exactly as practical as it sounds. He even worked with physicist Max Tegmark on making reinforcement learning interpretable, and contributed to research on autonomous lunar navigation in darkness. The Moon, then the funnel.
Then he did the thing that defines him more than any single paper: he left. With his MIT PhD nearly within reach, Mehta walked away from academia to build companies. Described by one profile as part mad scientist, part startup killer, he decided he would rather run experiments that talk back - customers, not citations.
BS, Computer Science & Applied Mathematics. Where the research habit started.
Master's in deep learning. Co-taught Duckietown - a course where students program tiny self-driving cars on a model town.
Computer vision at the Jet Propulsion Lab; robotics research in Seattle. Lunar navigation in darkness included.
PhD studying data distributions and multi-task learning - then a deliberate exit toward building.
His GitHub badges include “Arctic Code Vault Contributor” - his code sits archived in a vault in Svalbard, near the seed bank, for the next millennium.
He co-taught Duckietown, where students program miniature autonomous vehicles to navigate a model town. Robotics, but adorable.
He and his co-founder ran CharacterQuilt as a hands-on agency first - doing the grunt work by hand so they would know exactly what to automate.
A brain that learns your brand, and agents that operate the tools you already have.