Edward Wang Learned to Build Before the Camera Arrived

There is a small game hiding near the bottom of Edward Wang’s public project list. It is a 2D version of Super Mario, assembled from scratch with an agile team, event-driven programming and a patient supply of design patterns. In another résumé it would be charming debris from university: proof that someone once had coursework and free evenings. In Wang’s, it reads more like a clue. Long before he arrived at HeyGen, where faces, voices and language must agree inside a single frame, he was already interested in what happens when many moving pieces are forced to behave like one thing.

Wang is a founding software engineer at HeyGen in Los Angeles. The title places him near the machinery of an AI video company whose promise to customers is almost comic in its brevity: write a script and get a video. Simplicity on the screen, of course, is complexity wearing good tailoring. Behind the prompt are identity systems, model calls, rendering, timing, voice, lip synchronization, localization, storage and delivery. A user sees a button. An engineer sees a parliament.

The interesting part of Wang’s route is how many smaller parliaments he encountered first. Quantitative development at Finovax. AI software engineering at Lepu Medical Technology in Beijing. Teaching at Ohio State. Health technology at Cornell Tech. A software engineering internship at Microsoft. Two reported startup ventures, WZTrade and Lumo. The names change faster than the underlying instinct: take data, turn it into behavior, then build enough of the surrounding product to make that behavior useful.

“Beyond an engineer, I’m a curious quant trader, startup builder, and long-time crypto…”Edward Wang, in his public profile

The résumé as a workshop bench

Wang’s earliest listed professional work began in quantitative development in late 2021. Finance is a tidy place to learn an untidy truth: a model can be elegant and still be wrong by lunchtime. A few months later he moved into medical technology, working as an AI software engineer. Then came a teaching-assistant role at Ohio State. Taken together, the sequence put him in three different relationships with technical knowledge. In one, a model must confront a market. In another, software enters a specialized domain. In the third, understanding has to survive being explained to somebody else.

His student work was similarly unwilling to stay in one lane. For an ActivPlant project using Honda data, Wang helped turn raw manufacturing information into inputs for discrete-event simulation. The team built regression models intended to anticipate downtime and interactive graphs meant to make the output legible. The point was not merely to predict. It was to give planning a better instrument.

A pediatric sleep-state classification project went in the opposite direction, deeper into the model. Wang worked with accelerometer data, engineered time- and frequency-domain features, and used more than 1,500 of them during training. The project compared Random Forest, LightGBM and XGBoost, reporting an improvement of up to 32 percent and finding its strongest performance with XGBoost. It is a very student-like project in the best sense: ambitious, methodical and slightly intoxicated by the number of features available.

Another project analyzed public camera data for signs of potential violence, pairing YOLO object detection with speech recognition and text classification. The subject is serious, and the public description is technical rather than philosophical. Still, the combination matters. Video was not treated as a flat stream of pictures. It was a bundle of visual and audio signals that needed different tools before they could produce one judgment. Years later, Wang would work at a company whose product depends on a similarly multimodal bargain, though directed toward expression rather than detection.

A rocket, a checkout and the whole stack

If the machine-learning projects show Wang learning to choose among models, his full-stack projects show him refusing to stop at the model’s edge. For a NASA-themed launch-control system, he led backend work in Node.js for a React and Redux interface. The application used live SpaceX data, exposed REST APIs, added Google OAuth, tested with Jest, ran through a GitHub Actions pipeline, moved into Docker and landed on AWS EC2. NASA did not employ him, and the project was not a launch system used by the agency. Its value was the rehearsal: a miniature production environment in which every layer could object.

His e-commerce build performed the same exercise in a more familiar costume. A React storefront connected authentication, a database, application state, Stripe payments, tests and deployment. Shopping sites are excellent teachers because they punish vague thinking. A customer can be logged in or not. A cart can be current or stale. A payment can succeed or fail. Nothing is improved by calling the situation “mostly working.”

Then there is Mario. A game loop makes abstraction visible. Events have consequences; state changes; the screen must keep up. Wang’s team built the project from the ground up, and he described leading the work around patterns and interactive systems. It is not hard to draw a line from there to later product engineering. The character jumps only if input, logic and rendering agree. An AI avatar speaks convincingly only if a more complicated set of systems reaches the same compact agreement.

Health tech, startup weather and a large-company summer

From 2023 to 2025, Wang attended Cornell Tech and identified with the Jacobs Technion-Cornell dual master’s program in Health Tech. The campus is designed around technical work that expects to meet a market, a patient, an institution or all three. It was a fitting setting for someone whose project list already moved between prediction and product.

In the summer of 2024, he interned as a software engineer at Microsoft. Public records do not describe the team or product, which makes embellishment tempting and useless. What can be said is enough: after working in young ventures and academic projects, Wang spent a season inside one of software’s large institutions. Scale looks different there. A clever feature shares the room with compatibility, process, security and the accumulated memory of millions of users.

The same period included two founder chapters. Public career records identify Wang as co-founder and CEO of WZTrade beginning in 2024, and as co-founder and CEO of Lumo from September 2024 to March 2025. The available record says little about either company’s product. Their importance is biographical rather than promotional. Wang was not merely collecting engineering assignments; he was testing the other side of the table, where the questions include what to build, for whom and whether anybody returns tomorrow.

Early startups have their own weather system. Priorities arrive sideways. Product decisions become infrastructure decisions before lunch. A founder may switch from strategy to a broken integration without the ceremony of changing hats. That experience can be useful preparation for a founding-engineer role, where ownership tends to expand until it meets a wall, and then somebody asks whether the wall is load-bearing.

The camera disappears, the engineering does not

Wang joined HeyGen in 2025. The company had begun with a desire to make video creation less dependent on cameras, actors, locations and repeated shoots. Its product now spans avatars, translation, voice and interactive video across more than 175 languages and dialects. In June 2026, HeyGen announced more than 30 million users in 196 countries and $200 million in annual recurring revenue. Those are company figures, not Wang’s personal achievements. They describe the size of the room in which his engineering now operates.

Growth changes the character of a technical problem. At classroom scale, a bug disappoints the demo. At global scale, the same bug acquires time zones. Video adds its own appetite for computation and its own sensitivity to tiny errors. A mistimed mouth, a drifting voice or a face that feels almost right can undo a great deal of code. The product has to coordinate systems while protecting the human impression that no coordination was necessary.

This is where Wang’s broad route becomes more than résumé color. Quant work taught the suspicion that models deserve. Medical and health-tech projects required domain context. Teaching demanded explanation. Full-stack builds forced individual components into a usable whole. Startup work made technical judgment answer to a customer. Microsoft offered a view of engineering inside a large organization. HeyGen gathers those lessons in a product whose output is judged in seconds, often by people who will never know how many services had to cooperate.

The medium kept changing. The job stayed remarkably consistent: make complicated parts produce an intelligible result.YesPress

Wang’s career is still young, and the public record does not supply the grand confessions that profiles often use as scaffolding. There is no manifesto to pin above the desk. There is, instead, the evidence of repeated construction. A factory model. A classifier. A commerce site. A space-data backend. A game. Two ventures. An AI video platform.

That may be the more useful portrait. Engineers are often introduced through the finished thing, as if the title appeared at the same moment as the product. Wang’s story is about the rehearsals. Each project gave him another way to see a system: as prediction, interface, business, lesson or moving world. By the time the camera disappeared, he had spent years learning what had to remain behind it.