Alexzendor Misra and the business hidden inside a search engine

The interview ended, and Alexzendor Misra could not tell whether it had gone well. He and his co-founders had just spoken with Y Combinator’s Jared Friedman. They had applied three times and reached the interview twice. Then came an email with another Zoom link. On that next call, the uncertainty finally lifted: Shofo was in.

A founder can describe acceptance as destiny, or preserve the uncomfortable bit just before it. Misra preserved it. His account dwells on the gap between doing the work and knowing whether anyone else has been convinced. The photograph comes later. First there is an ordinary meeting, an ambiguous ending, and an inbox.

By then, he had already been CEO of Correkt, an AI search engine. At Shofo, he became co-founder and CTO. The change in title accompanied a change in customer: from people seeking answers to teams seeking material for training AI. Following that shift leads into the less photographed work of indexing, sorting and preparing information.

A search engine, with receipts

Misra arrived at UC Santa Barbara in fall 2022 to study computer science. A disagreement with a peer over conspiracy claims helped prompt the project that became Correkt. He assembled a student team through a Data Science UCSB competition; the project continued after the competition finished. The website launched on April 25, 2024.

The premise was straightforward: an AI answer should let its reader inspect where the information came from. Correkt paired generated responses with citations. The appeal was easy to understand on a campus, where an assertion and a reference have rather different powers of persuasion.

By May, the university had a place for that question on its calendar. Its AI Community of Practice Spring Symposium included a roundtable on platforms built for student success. Correkt’s presentation addressed ethical AI, credibility and accurate search. It shared the session with Storkie, the Student Affairs chatbot.

That setting matters. Before this was a story about supplying AI labs, it was a student project being discussed among people thinking about how AI belonged in university life. Credibility had practical consequences: could a student follow an answer back to something they could examine? A polished sentence alone was a fairly modest deliverable.

When the pictures arrived

Correkt subsequently announced another feature: images from the web would appear alongside text in its citation-backed answers. Misra responded publicly that images introduced another modality to AI search. The announcement put a visible example behind the technical word. An answer could now show something as well as describe it.

Pictures complicate the pleasant simplicity of a search box. Text has to be retrieved and related to a question; an image must also be relevant to what the answer is trying to explain. Adding a medium adds decisions. A user sees the result. The people building the product inherit the decisions underneath it.

His former company eventually reached 43,000 users, the figure in his founder biography. That supplies scale, but the more interesting continuity lies in the kind of work. Correkt brought material into an answer. Shofo prepares material that can go into a model’s training. In both cases, finding something is only useful if it fits the request.

An early Shofo announcement still described a product for discovering information. Co-founder Braiden Dishman introduced semantic social search, beginning with Reddit, with plans to add platforms and a customizable feed. Misra, Andre Braga and Bryan Hong were named alongside him. The invitation asked people to try it and offer blunt feedback.

That version of Shofo is a useful stop in the story. It prevents the later video business from looking as though it arrived fully formed. There was a search product, a request for feedback, and a stated interest in working with businesses that wanted the collected data. The audience was already widening.

The call after the interview

For Misra, the explanation for YC’s eventual yes was progress between interviews. He described nights spent indexing TikTok, X and Instagram, along with acquiring the first customers. The account gives acceptance a less ceremonial cause: the company had changed between one conversation and the next.

“It was our third time applying. Our second interview.”Alexzendor Misra, announcing Shofo’s YC acceptance

The distinction between applications and interviews is wonderfully specific. It keeps the repetitions visible. A badge compresses all that into two letters; his account puts some of the elapsed effort back. Shofo joined the Winter 2026 batch. Misra’s position was co-founder and CTO, with responsibility expressed in the language of a technical founder.

He illustrated the ambition with a very particular request: first-person cooking footage in Arabic, within a defined period, with another condition concerning the comments. It was an example of the precision the company wanted to support. The imagined customer was asking for footage that satisfied several requirements at once.

A person casually browsing cooking videos can tolerate a detour. A research dataset is defined by what belongs in it. Camera perspective, language and activity each change the selection. Pile enough conditions together and a familiar internet pastime becomes a retrieval problem. Misra’s example made the problem legible without requiring a lecture on infrastructure.

A biography in the metadata

Some of his work can be followed at a smaller scale than a launch announcement. The history of Shofo’s public TikTok dataset records contributions under AlexzendorMisra: uploading metadata, updating it, and changing paths to sharded files. These are practical entries about how someone else will find and load the material.

There is a pleasing reversal here. A founder’s title takes up very little space and can imply almost anything. A metadata change is narrow enough to explain. Files need addresses. Records need to point to the right files. The researcher who cannot load a video has little immediate use for a stirring mission statement.

The dataset includes fields for transcripts, descriptions, comments, engagement, language and technical properties such as frame rate and duration. Those fields show several ways of approaching the same clip. What someone says, what a creator writes beneath it, and how the file is encoded answer different questions.

Shofo’s English talking-head release makes the selection work particularly clear. Its description gives roughly 10,000 clips and 186 hours. A qualifying clip needs a centered, persistently visible face, active English speech, and checks connecting the voice to the face on screen. The release pairs footage with time-aligned transcripts.

Even this apparently modest category contains a small mountain of decisions. Someone speaking to a camera might have music underneath, an object across their face, or somebody else’s voice. A useful definition must deal with those possibilities. The dataset’s criteria turn the everyday phrase “talking head” into something an engineer can test.

The distinction offers a close view of the business Misra now helps build. An enormous collection is a supply. A collection selected for a purpose is a product. The labor of writing those requirements and applying them is easy to miss when everything arrives as a download.

The search box comes back

RecipeBook, Shofo’s self-service interface, brings the story back to a familiar piece of software: a search field. The interface offers ratings, filters, a control to train a probe, and an order measured in hours. A visitor can begin with a query and refine the selection rather than trying to describe every preference perfectly at the outset.

The vocabulary may have changed since Correkt, but the encounter remains recognizable. Someone wants information. A system returns candidates. The person needs a way to decide whether those candidates fit. Rating and filtering make that judgment part of the interaction. The search box has acquired a different destination.

Misra’s public technical interests include open-source language models, natural language processing, autonomous agent systems and generative AI. Those interests sit comfortably beside the work, without requiring a prediction about where his career will go. His stated ambition at Shofo concerns the material that future vision and multimodal models will learn from.

The current company website extends the catalog across computer use, talking-head footage, speech, audio and video games. Each comes with its own accompanying information: input logs for computer activity, transcripts for speech, recorded controls for gameplay. The useful unit can be an action together with the evidence needed to interpret it.

Misra’s story so far has moved from a campus argument to a search engine, then into the preparation of training data. The early concern was whether a reader could examine an answer. The present work asks what an AI team can examine and use before training begins. Between those questions sit repositories, filters, definitions and several attempts to get through an interview.

The next time someone types a simple request into a search field, a considerable amount of work will have gone into making the result look simple. Misra has spent this part of his career on that work. The photograph outside YC catches four people by a sign. The metadata offers another way to see what they have been doing.