Real-time AI characters using AI as the engineer

​Most teams building with AI agents hit the same wall. The agent burns half its context window loading environment tools on startup. Within a few hours, it starts re-discovering bugs you already fixed, re-asking for node IDs it memorized last week, and re-proposing architectures you already rejected.

​Mihai Satmarean's team at skylitetek.com ran 50+ continuous build sessions on a single project without that problem. Same recall as the linear context baseline. Dramatically fewer tokens. The memory layer in their workflow — Semvec, built by Versino PsiOmega GmbH (patent pending) — is why.

​The project itself is HoloTalks: live, cloud-streamed holograms you can walk up to and have a real-time conversation with. No generic chatbot masks. No client-side install. Sub-2-second response loops with synchronized lip movement and distinct character voices.

​Two characters are live: Leonardo da Vinci and Constantin Brâncuși. The agent handled the Unreal Engine Blueprint wiring, the C++ lip-sync patches, QA in Play mode, and AWS deployment. A traditional 3D agency delivered the static MetaHuman mesh. From there, the agent took over.

​This talk is the unfiltered engineering story. Mihai will walk through the architecture, the memory system, the mistakes, and what the actual bottlenecks look like when AI is your engineering team.

​Your Hosts:

​Vibhuti Ranjan Singh Founder & Managing Partner, Product Advisors

​Mihai Satmarean Technology Partner, Product Advisors | Co-Founder, Skylite Tek GmbH

​Tashi Yangdhar Head of BD & Partnership, Product Advisors

​By registering, you acknowledge that this webinar will be recorded and might be use for educational and marketing purposes.