Yoeven Khemlani had the exit and the beach. Then a question he couldn't answer - why can AI talk but not work? - turned into JigsawStack, then into a Y Combinator company called Interfaze.
The plan was to stop. Yoeven Khemlani had done the hard thing once already - co-founded a travel startup in Singapore, carried it through the worst years the hospitality business had ever seen, scaled it across Southeast Asia, and sold it. That was supposed to be the whole story. Exit, breathe, disappear. So he moved to Bali and set out to do exactly nothing.
It did not take. The problem with retiring from building things is that you have to stop noticing the things that are broken, and Khemlani has never been very good at that. The thing he kept noticing, from a beach that was meant to be the end of his working life, was a gap that most people were happy to talk around: the new wave of AI could hold a conversation, write a poem, riff on almost anything - but ask it to behave like a software engineer, to read a document's layout, pull structured data out of a mess, and return the same reliable answer every single time, and it fell apart.
AI could talk. It couldn't work. And the difference between those two things is where every real business lives.
The problem that ended a retirementThat distinction - talking versus working - is the whole of Khemlani's second act. He is a builder by temperament, the kind who reaches for a laptop the way other people reach for a notebook, and eventually the itch won. He opened the machine back up and started testing.
He did not start with a manifesto. He started with numbers, feeding real-world tasks - the unglamorous kind, reading actual documents and live websites - into whatever models he could get his hands on and writing down what came back. The results were a ladder, and each rung told him the ceiling was higher than the one before.
Off-the-shelf models topped out fast. A model built for the job got to 80% on live websites - the number that convinced him there was a company here.
Twenty percent from GPT-3. Fifty from a fine-tuned Llama. Eighty from something he trained himself, aimed squarely at the task. Most people would have quit at the first rung and concluded the technology wasn't ready. Khemlani read the same ladder and saw the opposite - not a dead end, but a slope worth climbing, if you were willing to stop treating a general-purpose chatbot as the answer to a specialist's problem.
The insight underneath the numbers is quieter than most AI pitches, and it's the part worth stealing. Khemlani decided the scarce resource wasn't intelligence. It was reliability. A model that dazzles nine times out of ten is a party trick. A model that has to be right the ten-thousandth time, unwatched, in a banking pipeline or a KYC check or a logistics system, is infrastructure - and in those places a single wrong character doesn't dent the output, it breaks the whole workflow.
Enterprises don't need an AI that's brilliant most of the time. They need one that's boring and correct every time.
The bet behind InterfazeSo he built for correctness. The architecture he landed on is a hybrid: traditional deep-learning vision models - the CNNs and DNNs that come out of years of computer-vision work - handling what they're best at, paired with transformers doing the reasoning on top. The trick is that the two halves check each other. Where the vision layer misreads, the reasoning layer can catch it; where the reasoning layer hallucinates, the vision layer holds it to what's actually on the page. Two different kinds of machine, each covering the other's blind spot.
The prototype became JigsawStack, a suite of small, custom-built models packaged as a single API - web scraping, visual OCR, translation, speech-to-text, classification - the plumbing that developers usually stitch together from a dozen fragile services. It found an audience and it found backers: roughly $1.5 million in pre-seed money from investors including Antler and Ada Ventures, whose deal was filed under a theme they call economic empowerment.
Growth brought its own problem, the kind founders are lucky to have. As usage climbed, GPU costs threatened to swallow the whole thing. The fix came from specialized inference hardware built for real-time speed, which made the economics work at scale. That unlock is what let the idea widen into Interfaze - the same obsession, bigger swing - stitching OCR, scraping tools and workflows into one dependable system meant to run at volume.
Interfaze's own claim is blunt: on the deterministic tasks it targets, it aims to beat the general-purpose heavyweights - Gemini Flash, Claude Sonnet, GPT-4 mini - not on charm, but on getting the same answer right, again and again. It ships with the parts developers actually ask for when they're building for production rather than for a demo: confidence scores, bounding boxes, verifiable extractions, support for a hundred-plus languages, and an API that speaks the same dialect as the tools teams already use. In 2026 it joined Y Combinator's Spring batch. Khemlani had also done the thing a lot of Asian founders were doing that year: packed up from Singapore and moved to the San Francisco Bay Area, closer to the customers who buy this kind of infrastructure.
It's worth sitting with how contrarian the timing was. The rest of the industry spent that stretch in an arms race over size - more parameters, longer context, bigger everything - on the theory that a sufficiently large model would eventually be good at every job. Khemlani pointed the other way. His argument, made on podcasts and in the product itself, is that a fleet of small models each trained for one task will beat a single giant one on the tasks that matter to a business, and do it cheaper and faster. It is the kind of position that sounds obvious only after someone has been right about it.
Read his back catalogue and the AI founder starts to make more sense. Before any of this, Khemlani was a game developer - Unity, C#, the psychology of why a mechanic feels good. His portfolio is a wonderfully odd list: a hotel-booking app, a customer-journey tool for Singapore Airlines, a cybersecurity training game called "What the Hack!", and a Kinect-based system that used motion capture to teach Thai dance. On GitHub, where his bio reads simply "Building to my heart's content," his most-starred project is an AI video search engine that pulls in six hundred-plus stars.
A decade of building, from motion-capture dance games to AI infrastructure - the throughline is fixing what annoys him until it works.
The game-design instinct shows up in the AI work more than it looks. Games are systems that have to behave consistently or the illusion collapses; a physics engine that's right most of the time is a broken game. It's not a stretch to see the same demand for consistency running underneath someone who now sells reliability as a product. The through-line across a decade of very different projects is a builder who fixes what annoys him and doesn't much care whether the fix is fashionable.
That temperament also explains the shape of his companies. Both JigsawStack and Interfaze are, at bottom, the same instinct expressed twice - take the tedious backend work that developers dread wiring together, and turn it into something you can call in a single line. He is not chasing the flashiest corner of AI, the assistants and agents that make the headlines. He's chasing the part that everything else quietly depends on, which is a less glamorous place to plant a flag and, historically, a more durable one.
On his personal site, Khemlani keeps a line from the biochemist Albert Szent-Gyorgyi: "Innovation is seeing what everybody has seen and thinking what nobody has thought." It's a tidy summary of his whole move. He looked at the same large language models everyone else was looking at, watched them dazzle in demos and stumble in production, and thought the un-obvious thought - that smaller and narrower and less impressive might actually be the smarter bet, precisely because it could be trusted.
Innovation is seeing what everybody has seen and thinking what nobody has thought.
Yoeven Khemlani's guiding lineThere's a version of the founder story that ends in Bali, on that beach, with the money made and the ambition spent. Khemlani got about as close to it as anyone does and then quietly turned around, because the problem was more interesting than the rest. He is building, again, to his heart's content - this time on the unglamorous, load-bearing part of AI, the part that has to work.