Ondrej Texler, Principal Research Scientist at ValkaAI – iGaming Interview Series

Ondrej Texler is Principal Research Scientist at ValkaAI, with a background in computer graphics, artistic stylization, and photorealistic virtual humans. His research has spanned tools for transferring visual styles to images and video, as well as avatar technology at Samsung and HeyGen. He now works on neural rendering and controllable digital characters, including the technology behind Oddin Casino’s AI-powered digital dealer.

ValkaAI develops real-time AI video technology and lifelike digital personas for entertainment. Its work combines neural rendering, generative audio, and motion modeling to create characters that can respond to context and interact with audiences. This interview explores how those capabilities apply to digital dealers, from coordinating speech and movement to separating generative interactions from deterministic game mechanics.

Your career has taken you from developing algorithms to help artists to transfer painterly styles to images and creating animated stylized videos to researching photorealistic virtual humans at Samsung and HeyGen. How have these experiences shaped your approach to developing Oddin Casino's AI-powered digital dealer?

The unifying thread throughout my career has always been control. In my style transfer research, the goal was to give artists complete command over the generation process—ensuring every brushstroke, canvas texture, and color palette strictly matched the example they provided. When I moved to photorealistic virtual humans at Samsung and HeyGen, that same philosophy applied to digital humans: we needed surgical control over micro-expressions, gaze direction, and phoneme-accurate lip sync without any hallucination so that the person’s identity is perfectly preserved.

For Oddin Casino’s AI digital dealer, this exact level of steerability is everything. My past work directly informs how we build our neural models to be strictly controllable—allowing our digital dealers to execute precise table etiquette, adapt instantaneously to player bets and game states, and maintain authentic, natural mannerisms under real-time constraints.

Oddin Casino is developing its AI dealer using proprietary models rather than combining a conventional game engine with third-party AI services. What are the key technical advantages of building this technology at the model level, and what capabilities does this approach make possible?

We decided to take the hard path because we believe that’s the correct path, and I believe it is the only successful path if we really are aiming for a real-time, immersive, and truly photorealistic experience. If we used any traditional game engines, it would be impossible to achieve fully photorealistic, “camera-like” quality.

There are games out there that have some of the most amazing and realistic graphics that the industry has seen, but you can tell within seconds that it’s a 3D game, not camera footage. For us to get to the level of immersiveness that we’re after, customers should not be able to tell if the dealer is live or AI-generated. A video model is the only way to achieve that.

Creating an AI character that can speak, react, and interact naturally in real time presents significant engineering challenges. How does your system coordinate speech generation, facial expressions, body movements, and player interactions while maintaining the low latency necessary for a seamless experience?

Exactly, this is the ultimate challenge that needs significant engineering and research effort. I’d split it into two categories. First, individual pieces: the models themselves need to run in real time. And there are a lot of such individual models: text generation, speech generation, body movement, card movement, and finally a photorealistic video rendering model. Some models have tens of billions of parameters, and optimizing them requires a significant research effort. Second, the entire inference pipeline that runs all the models needs to be designed to allow for a very fast response when a player action comes in. It took us over a year to come up with a framework and set of models that deliver satisfactory quality and a real-time experience, and despite all optimizations, it still needs a lot of very beefy GPUs to run.

One of the biggest challenges with AI-generated characters is overcoming the uncanny valley and making interactions feel genuinely natural. What breakthroughs in neural rendering, motion modeling, or generative AI have been most important in achieving realistic digital dealers?

I do not think there is any single breakthrough; it is a series of continuous improvements. Over the last couple of years, the capability of image and video models has advanced beyond our wildest imagination, and we are slowly reaching the point where the uncanny valley is not really noticeable on the visual or textural level.

Some 4-6 years ago, digital humans looked and talked absurdly bad; that’s not the case anymore. However, the uncanny valley still remains on the behavioral level, e.g., digital humans often having a “weird” look, reacting poorly to certain situations, having “unnatural” motions and eye contact. It will probably take a few more years before real-time digital humans are truly unrecognizable, especially when you can interact with them in real time.

Oddin Casino's development has reportedly involved more than two million GPU hours of training. What have been the most computationally demanding aspects of developing these models, and how do you balance visual realism, responsiveness, and computational efficiency?

Learning coherent movement across video sequences is particularly demanding. The models need to preserve identity and visual detail while handling expressions, hands, and interactions with objects. The training investment behind this technology also includes experimentation: testing approaches, identifying weaknesses, and refining the models. As researchers, we have to evaluate visual quality alongside responsiveness and inference cost. A beautiful result only becomes useful when it can be delivered quickly, reliably, and economically.

Oddin Casino describes its technology as capable of adapting a dealer's personality and remembering previous player interactions. How does the underlying AI maintain a consistent personality while responding dynamically to different users, and what safeguards govern personalization and player data?

I see personality as a stable identity with flexible expression. A dealer can remain the same recognizable character while adjusting its pace, humor, or level of conversation. Memory provides relevant context rather than changing that identity. My principle is that persistent memory should have a clear purpose, appropriate consent, and defined access, retention, and deletion controls. Personalization should make interactions more comfortable, without using player vulnerability to encourage further gambling.

Unlike conventional gaming NPCs, an AI dealer operates within an environment where predictable behavior, game integrity, and accountability are essential. How do you separate generative AI interactions from deterministic game mechanics, and what safeguards prevent incorrect or inappropriate responses?

The separation must be explicit: the game system determines rules, outcomes, and settlement; the AI character communicates and interacts. Oddin Casino describes outcomes as coming from certified game logic, with the avatar having no knowledge of the next card. From a technical perspective, game-related statements need validation, interactions need to be reviewable, and uncertainty needs a controlled fallback. A conversational error must never become a game-state error.

Your technology is designed to allow operators to customize digital dealers for different brands, languages, and markets. How much of this customization can be achieved without retraining the underlying models, and what technical challenges arise when adapting personalities, voices, and behaviors across cultures?

Many aspects, including conversational style, brand guidelines, and language selection, can be configured without retraining the underlying models. Changes to appearance or voice may require additional assets or adaptation. The harder challenge is maintaining a coherent character across markets. Localization goes beyond translating words: humor, politeness, gestures, and conversational timing all vary. Native-language and cultural evaluation are essential, because a fluent character can still behave in ways that feel unnatural.

Running sophisticated generative AI models continuously for potentially thousands of simultaneous users introduces substantial infrastructure challenges. How are you approaching inference optimization, GPU utilization, and scalability to make real-time AI characters commercially viable?

For the underlying technology, the key is to optimize both the models and algorithms and then run efficient inference. When developing the models, we are always trying to see what is the minimal computational complexity that we need to achieve the desired quality. We are carefully reviewing and evaluating how big a model architecture we need, what resolution we need, how many diffusion denoising steps we need, and countless other things. But let’s face it, to maintain high quality, even with all of the optimizations, the computational complexity is significant, and so is the inference cost. From the product and economic viability point of view, it matters whether the system is serving a shared table or individual players: thousands of viewers do not necessarily require thousands of independent tables–generations.

Looking beyond iGaming, advances in real-time AI characters could have significant implications for video games, virtual worlds, esports broadcasting, and interactive entertainment. Do you envision this technology eventually powering a new generation of intelligent NPCs and digital personalities that can interact with audiences in ways traditional scripted characters cannot?

Absolutely. Digital dealers are a focused application of a broader capability: characters that understand context, communicate naturally, and maintain a recognizable identity. In games, that could mean NPCs that respond meaningfully to a player. In broadcasting, it could mean digital hosts that interact with audiences as events unfold. A use case I used to work on that still excites me is creating tools and frameworks for artists and content creators, and the technology we are developing certainly has potential in this domain as well.

Thank you for the great interview, readers who wish to explore real-time AI characters, controllable neural rendering, and digital dealer technology should visit ValkaAI.