The generative AI space is crowded with tools that promise to revolutionize how we create and consume video content, but few challenge the very nature of what video can be. Eqvista sits down with Milos Lokajicek, Co-Founder and COO of ValkaAI, a deep-tech startup that is pushing beyond today’s AI-generated media.
While most of the industry is building better photocopiers, ValkaAI is building a different engine. Their real-time interactive video stack treats video as an interface, not just an output, with avatars that interact with their 3D environments in milliseconds, aiming to move the industry from passive “content on demand” to interactive “experiences on demand.”
Built as a global company from day one, with hubs in the US and Prague, the team is a deliberate collision of two worlds: academic heavyweights pushing the boundaries of neural networks, and seasoned operators who know how to ship production-grade code. In this interview, Milos shares his candid insights on everything from technical bottlenecks and model governance to hiring strategy and investor relations.
How would you position ValkaAI in the crowded generative AI and digital human landscape—what makes your real-time interactive video stack fundamentally different?
Most of the industry is currently building better photocopiers—tools that generate linear, pre-rendered, static video clips based on text prompts. ValkaAI is building a fundamentally different engine. We are developing a real-time interactive video stack where video is an interface, not just an output. Our avatars won’t just speak; they will dynamically interact with their 3D environments in milliseconds. We are aiming to move the industry from passive “content on demand” to interactive “experiences on demand.”
What’s the key technical bottleneck in real-time AI avatars, and how is ValkaAI solving it?
The massive bottleneck is balancing latency with temporal consistency. When standard diffusion models generate long sequences, errors multiply over time, leading to flickering, melting, or hallucinated limbs.
Our research team solved this by developing a proprietary control model and a “language of motion.” Instead of generating heavy, continuous chunks of video that degrade, we generate frames independently based on real-time inputs. This maintains perfect visual consistency and allows for hyper-personalized, unscripted movement without the massive latency overhead.
From an operating-model perspective, how do you balance R&D intensity with productization and commercial deployment?
You cannot do deep tech R&D in a vacuum, or you will just burn capital. We anchor our massive R&D efforts to highly constrained, immediate commercial use cases. For example, our first major product is an AI commentator for live sports and esports. By focusing on this specific vertical, we get our product into the market fast, and the real-world data from these broadcasts feeds directly back into our core research engine.
How do you approach talent and team structure for a highly specialized AI-avatar startup?
We built ValkaAI as a global company from day one, with hubs in the US, and Prague. Our team structure is a deliberate collision of two worlds: academic heavyweights pushing the boundaries of neural networks, and seasoned operators who know how to ship production-grade code. You need hardcore distributed-systems engineers to make heavy AI models run in real-time.
How do you approach model governance and safety for AI personas in sensitive verticals like news and sports?
When a digital persona makes a mistake on a live broadcast, it’s not a glitch—it’s a potential PR disaster for the network. That’s why we absolutely do not let our avatars “think” freely using off-the-shelf, open-ended LLMs. Instead, we rely on highly restricted cognitive architectures. We tether the AI’s brain directly to live, verified data APIs—like real-time match telemetry. The avatar’s job isn’t to invent commentary; it’s to stylize and verbalize hard facts with the right emotional cadence. By stripping away creative freedom where it doesn’t belong, we effectively eliminate the risk of hallucination.
From your perspective, how should founders think about regulation, identity, and watermarking before they hit trouble with broadcasters or sports leagues?
Founders must be proactive. Broadcasters have zero tolerance for deepfake risks or IP infringement. If you wait for a cease-and-desist, you’ve already lost. You must bake provenance and watermarking into your architecture from day one. In enterprise media, trust, clear IP rights, and transparent AI signaling are just as important as frame rates.
Congratulations for raising 12M in the pre-seed round. What KPIs are tied to post-funding milestones, and how do you report them to your investors?
Raising a €12M pre-seed is an incredible vote of confidence, but it immediately starts the clock. Because we are building fundamental deep tech, our board isn’t hunting for standard software metrics like MRR just yet. They are looking for hardcore technical de-risking. Our KPIs are almost entirely milestone-based: hitting strict sub-second latency thresholds at scale, achieving key breakthroughs in our motion-rendering pipeline, and successfully executing high-stakes commercial pilots. We keep our reporting incredibly transparent around these engineering benchmarks, because solving the core tech is the only way we unlock this entirely new category.
After closing a large pre-seed round, what’s your advice on hiring fast without destroying culture or over-engineering the org structure too early?
Don’t hire just because you have the cash, and do not build a massive corporate org chart. Hire “T-shaped” individuals who thrive in ambiguity. Keep the hierarchy incredibly flat but the goals razor-sharp. Our leadership is deeply involved in onboarding to ensure every new hire is obsessed with our core mission. Over-engineer your communication, not your management layers.
From an investor or corporate-innovation perspective, what should stakeholders watch in the “AI-human” space over the next 3–5 years?
Watch the death of passive streaming. The era of “prompt and wait” is ending. Stakeholders should look for the companies solving the “uncanny valley of interaction”—which isn’t just about photorealism, but conversational pacing, interruption handling, emotional intelligence, and real-time contextual awareness.
For first-time AI founders, what’s the one thing you wish you’d known before starting ValkaAI about balancing R&D with product-market fit?
Because ValkaAI isn’t my first startup, I had the advantage of learning from my past mistakes in other projects. And the biggest lesson is that deep tech without a tight, brutal feedback loop is a black hole. It is incredibly tempting for founders to fall in love with their own research and spend years in stealth building a “general-purpose” AI that nobody actually knows how to buy. My advice? Pick one incredibly painful, hyper-narrow problem for a specific user, and over-optimize your tech exclusively for that experience. Get it into production fast. You have to let the messy friction of the real market dictate your R&D roadmap, rather than letting the lab dictate the product.
