Kaushik Mahorker and the work behind an answer

A croissant is an unusual sales tool. It has no analytics dashboard, no integration guide, and a regrettably short shelf life. Kaushik Mahorker nevertheless tried it. After unanswered outreach, he described collecting pastries from Arsicault and taking them to brands around San Francisco. Some recipients accepted the food and sent him away. Others talked him through their processes. His account included a parking ticket. Customer discovery, apparently, has incidental expenses.

The episode offers an apt introduction to the co-founder and CEO of Wildcard. He has worked in cloud engineering and managed an AI engineering team. Now he is trying to understand how retailers operate when software helps decide what a shopper sees. There is code involved, certainly. There are also people who know their products, imperfect catalogs, and conversations that have to happen before an engineer can choose what to build.

Before the catalog, a virtual world

In June 2019, Mahorker was a fourth-year computer-science student presenting research at UC Santa Barbara. His project concerned creating customizable, textured virtual-reality environments from real-world images. Tobias Hollerer was his adviser. The university’s undergraduate poster session brought student work before faculty, fellow students, and the public. It was a setting for explaining an experiment, rather than announcing a company.

Mahorker also received Distinction in the Major on the department’s research track. That detail gives his early career a particular texture: there was an interest in investigating a technical problem and presenting the result. His later education included a master’s degree in computer science at UCLA. The degrees establish training; the projects and writing reveal some of the questions occupying that training.

In May 2020, he published an explanation of EWISE, a word-sense disambiguation approach developed by other researchers. His subject was a familiar nuisance of language: the same word can mean different things in different sentences. A system has to use context to choose the appropriate meaning. His essay walked through the architecture and its ability to handle unfamiliar or rare words.

For a future founder working with shopping conversations, it makes an interesting earlier subject. A request is more than a string of keywords. The surrounding words change what a person is asking for. Mahorker’s essay was an explanation of existing research, rather than a claim to its invention. That distinction matters. Teaching a difficult idea clearly is work of its own.

Three months later, his NoScope essay addressed another practical question: what happens when a neural network must process video at scale? He used a traffic-analysis scenario to explain specialized models and ways to avoid unnecessary computation. The attraction was efficiency: retain useful results while reducing the work needed to produce them. A demonstration becomes a different proposition when the volume increases.

Read together, the essays supply an early view of his technical interests. One concerns interpreting an input correctly. The other concerns making repeated interpretation affordable. Neither requires a shopping cart. Both belong to the less theatrical part of software, where something has to keep working after the first impressive result.

Four hundred thousand reasons to care about details

Mahorker subsequently worked on AWS Elastic File System and at Scale AI. At Scale, he became an engineering manager leading the GenAI Allocation team. These roles put his career in the territory of production systems and engineering responsibility. A founder biography can turn employers into decorative badges. Here, the specific work is more useful than the badges.

His account of a Scale ecommerce pilot describes an enrichment engine spanning 400,000 stock-keeping units and 2.4 million product attributes. A SKU identifies a particular item or variant; an attribute records something about it. At that volume, a catalog becomes a systems problem. Individual descriptions have to participate in a structure that can be processed repeatedly.

The number gives the later retail turn a concrete antecedent. He had already worked on organizing product information in quantity. That does not make the commercial choices of a new company automatic. It does mean that the product catalog was familiar technical terrain before it became the center of Wildcard’s current proposition.

The company changed its question

Wildcard began with a different public offering. In February 2025, Mahorker and co-founder Yagnya Patel launched a gateway for AI agents to find and use APIs. They had met as undergraduates. Patel’s earlier work included natural-language processing and knowledge graphs at Tesla, Amazon, and Truveta. The two brought different engineering experiences to a shared company.

The early Wildcard product included agents.json specifications and an open-source software development kit. Its task was to help an agent choose the API actions needed for a request and execute them on the user’s infrastructure. The launch described a collection of ten curated APIs. Wildcard was part of Y Combinator’s Winter 2025 batch.

That original problem was about translating a desired outcome into usable software actions. It is easy to see how it attracted engineers: an agent’s request might require several steps, while an API exposes particular operations. Something has to arrange the handoff. Wildcard’s first chapter took that gap as its subject.

By July 2026, Mahorker was explicit about the change. Wildcard had relaunched in November 2025 around agentic commerce. The company was no longer actively maintaining or working on the earlier protocol. Preserving that turn makes the story more intelligible. A company name can survive a substantial change in what its founders decide to sell.

He described the new ambition as helping ecommerce brands get their products discovered and purchased through shopping assistants such as ChatGPT and Gemini. At the time, he said Wildcard was working with more than twenty brands and preparing a broader rollout. Its scope included visibility tracking, product-page optimization, content work, and revenue attribution. The company linked its direction to seven months of conversations with brand owners.

The buyer does not speak in database fields

Mahorker’s published commerce guidance begins with a shopper’s constraints. A person looking for an air fryer may care about the number of people it serves, the counter space it occupies, and how easily its parts can be cleaned. Those requirements have to meet facts about the product. An attractive category name alone has a limited vocabulary.

His Walmart Sparky guide makes the operational implication specific. Dimensions, capacity, variant labels, and fulfillment details need to agree. A location can affect availability. A feed may overwrite a one-off edit. The recommendation is to correct verified information at its maintained origin and review repeated observations. There is considerable retail housekeeping behind a conversational answer.

This is also where Mahorker’s subject becomes recognizably human. Someone owns the item record. Someone understands the difference between two variants. Someone approves a claim before it goes live. A system can help organize the work, but it needs those decisions. A confident sentence is a poor substitute for a correct measurement.

“why not me?”Kaushik Mahorker, reflecting on starting his own company

In his broader agentic-commerce framework, the assistant may help a customer discover, compare, and sometimes purchase a product. The website still has a role in explaining and validating the item. Mahorker describes work crossing growth, merchandising, catalog operations, content, and engineering. The channel owner needs enough authority to turn an observed gap into a task with an owner.

That is a revealing choice of emphasis. The glamorous part of an assistant is its fluent response. His guidance spends time on who maintains the facts beneath it. It treats the catalog as shared operating infrastructure, with improvements useful across storefronts, feeds, marketplaces, and newer shopping interfaces.

A baseline before a victory lap

Mahorker’s 2026 operating plan for commerce leaders proposes a deliberately bounded start: one category, twenty priority SKUs, ten buyer questions, and two relevant shopping surfaces. The first thirty days establish a baseline. The next thirty focus on shipping corrections. The final thirty revisit the observations and decide what to expand, revise, or stop.

His guidance also distinguishes an observed mention from a universal ranking and attributable orders from less certain influence. Responses can vary with wording and context. A before-and-after result alone does not establish that a catalog edit caused a revenue change. Those limits are part of the measurement method he recommends, rather than an invitation to abandon measurement.

Another essay, responding to a 9operators discussion of personalized shopping, considers what happens when an assistant sits between a brand and a customer. Mahorker’s interpretation keeps existing direct channels in view. He argues for preparation and diversification: improve the information that travels beyond the storefront while continuing to support the routes that already work.

That brings his career’s practical thread into focus. The student explained research. The engineer worked on systems and large quantities of product data. The founder changed the company’s direction and began selling into retail operations. Each stage supplies a different kind of work behind the apparent simplicity of an answer.

Wildcard’s stated ambition is to connect observation, approved action, and commercial outcomes. For Mahorker, the next chapter lies in making that sequence useful to the people responsible for real products. A shopper’s request may arrive as a sentence. Somewhere behind the sentence, a product needs its dimensions, a task needs its owner, and a founder needs another conversation.