Tryll Engine

Waiting until everyone else finished posting their GDC Festival of Gaming pics before sharing ours, definitely a tactical move and not just us being late. Big respect to the Tryll Engine ambassadors, Andrei Morozov, Lidia Kozlova, and Pierre Moisan for being on the front lines and showing the industry what we’re building. Coming back, we’re honestly surprised by how much the vibe has shifted. AI optimism has skyrocketed in just the last few months. People are moving past simple experiments and looking for real, production-ready tech. This year is going to be massive for innovation. If you’re already adding on-device AI to your game, planning to add it, or better yet, want to work on it yourself-hit us up!

Tryll Assistant has been updated to version 0.0.10. This release brings a smoother onboarding flow, a cleaner interface, smarter dialogue handling, and support for 9 new games. What’s new in this update: - New 5-step setup wizard for first launch - Dialogue memory for smarter follow-up answers - Real-time model download progress on the home screen - Updated backend selection with clearer cards and tags - Improved chat experience with a dedicated send button - Redesigned Contacts, FAQ, and About pages - Better stability, logging, and bug fixes Download and try Tryll Assistant here: https://lnkd.in/dQvipHYy

GDC Festival of Gaming 𝗵𝗮𝘀 𝗼𝗳𝗳𝗶𝗰𝗶𝗮𝗹𝗹𝘆 𝘀𝘁𝗮𝗿𝘁𝗲𝗱 - 𝗹𝗲𝘁’𝘀 𝗴𝗼! 🚀  This week the game industry gathers in San Francisco, and the Tryll Engine team is on the ground. If you’re at GDC and want to talk about 𝗔𝗜 𝗶𝗻 𝗴𝗮𝗺𝗲𝘀, or the future of game development - book a meeting with us. Our team members Lidia Kozlova🔜GDC, Pierre Moisan, and Andrei Morozov are there all week, meeting studios, developers, and partners. You can find us on MeetToMatch! Always happy to connect, exchange ideas, and explore collaborations. See you at GDC! 

𝗪𝗲’𝗿𝗲 𝗲𝘅𝗰𝗶𝘁𝗲𝗱 𝘁𝗼 𝘄𝗲𝗹𝗰𝗼𝗺𝗲 Pierre Moisan 𝘁𝗼 𝗧𝗿𝘆𝗹𝗹 𝗘𝗻𝗴𝗶𝗻𝗲! Tryll Engine brings 𝗼𝗻-𝗱𝗲𝘃𝗶𝗰𝗲 𝗔𝗜 into games, enabling developers to run AI models directly on players’ hardware and create new kinds of gaming experiences. To help us build this future, we’re thrilled to be joined by Pierre Moisan, 𝗮 𝗽𝗶𝗼𝗻𝗲𝗲𝗿 of the Canadian video game industry with 𝟯𝟬 𝘆𝗲𝗮𝗿𝘀 of experience building and scaling game studios. As Tryll Engine continues to grow, Pierre will support us with strategic guidance, product insight, and deep industry connections across the North American game ecosystem. Earlier in his career, Pierre served as Vice-President at Megatoon, the studio that later became Behaviour Interactive today one of Canada’s largest game companies. He later founded ArtMedia Studios, and joined Ubisoft to help launch the company’s Quebec City studio. Pierre then spent nine years at Frima, where he was Vice-President of Strategic & Business Affairs, helping grow the company from a small team into the largest independent game studio in Canada. Today, Pierre is 𝗦𝗲𝗻𝗶𝗼𝗿 𝗔𝗱𝘃𝗶𝘀𝗼𝗿 for Creative Industries at CEIM in Montreal. As we bring on-device AI to the next generation of games, Pierre’s experience will be invaluable. Welcome to the team, Pierre !

Tryll Assistant v0.0.8 - 𝗖𝗨𝗗𝗔 & 𝗛𝗜𝗣 𝗦𝘂𝗽𝗽𝗼𝗿𝘁 In addition to the existing Vulkan backend, v0.0.8 introduces: - 𝗡𝗩𝗜𝗗𝗜𝗔 𝗖𝗨𝗗𝗔 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 - 𝗔𝗠𝗗 𝗛𝗜𝗣 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 Users can now choose the backend that best matches their GPU architecture CUDA and HIP backends are distributed as 𝗼𝗽𝘁𝗶𝗼𝗻𝗮𝗹 𝗗𝗟𝗖𝘀 𝗼𝗻 𝗦𝘁𝗲𝗮𝗺. To enable them, simply download the corresponding DLC - the backend selection will then become available inside the application settings. 𝗧𝗿𝘆𝗹𝗹 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 𝗶𝘀 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝗼𝗻 𝗦𝘁𝗲𝗮𝗺. Check it out - link in the comments.

We’re excited to welcome Alexander Stojanovic as an Advisor to Tryll Engine. Alexander is Vice President of NEXT AI R&D at Microsoft GenAI. Previously, he served as Managing Director of Gen AI/ML at Splunk and as VP / CTO of ML & AI at eBay, Cisco, and Oracle and more. We first met in Belgrade in October, and since then have stayed in close contact. Over time, Alexander has been supporting us with product insight and strategic advice. Today, we’re glad to announce that he is joining as an Advisor. Welcome to Tryll Engine

𝗧𝗿𝘆𝗹𝗹 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 𝗶𝘀 𝗹𝗶𝘃𝗲 𝗼𝗻 𝗦𝘁𝗲𝗮𝗺 It’s an 𝗼𝗻-𝗱𝗲𝘃𝗶𝗰𝗲 gaming assistant that runs locally on your PC. Press a hotkey in-game, ask your question, get a clear answer, and continue playing. 𝗜𝗻𝘀𝘁𝗮𝗹𝗹 𝗻𝗼𝘄 𝗼𝗻 𝗦𝘁𝗲𝗮𝗺: https://lnkd.in/dQvipHYy 𝗦𝗵𝗮𝗿𝗲 𝘆𝗼𝘂𝗿 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗼𝗻 𝗗𝗶𝘀𝗰𝗼𝗿𝗱: https://lnkd.in/dnweUpGC

𝗛𝗼𝘄 𝗪𝗲 𝗖𝗮𝘁𝗰𝗵 𝗣𝗿𝗼𝗺𝗽𝘁 𝗜𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻𝘀 𝗮𝗻𝗱 𝗛𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻𝘀 "Forget all previous instructions", "you are DAN now", "developer mode ON" this is not theory, these are real requests. We are building 𝗧𝗿𝘆𝗹𝗹 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 for gamers, and it works directly inside the game. From day one, we built security into the architecture. Here is how our pipeline works: User Input → Injection Filter → System Prompt → LLM + RAG → Semantic Filter → Response 𝗧𝗵𝗿𝗲𝗲 𝗹𝗲𝘃𝗲𝗹𝘀. Each one catches something different. 𝟭. 𝗜𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻 𝗙𝗶𝗹𝘁𝗲𝗿 - we catch manipulation at the input. 43 regex patterns detect prompt injection attempts before the request reaches the model. "Ignore previous instructions", "god mode", "safety is disabled", role attacks through characters - all this is blocked at the first barrier. If injection is detected, the model is not called at all. The user receives one of 20 neutral responses. 𝟮. 𝗥𝗔𝗚 - the model answers only about the game. The assistant doesn't just answer - it searches. Every answer is built on the knowledge base of a specific game: mechanics, NPCs, locations. If there is no relevant information in the database, the model will say so instead of confidently hallucinating. This is a key decision: better an honest "I don’t know" than a beautiful hallucination. 𝟯. 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗙𝗶𝗹𝘁𝗲𝗿 - we catch hallucinations at the output. Even with RAG, the model can go off topic. So at the output, we compare the answer with the original chunks using cosine similarity. If the answer is too far from the knowledge base, it will not pass. This is not post-moderation. It is a mathematical check of how close the answer is to real data in the knowledge base. What this gives in practice: - The user cannot force the assistant to leave its role - The model is much less likely to invent game mechanics that do not exist - Every answer is supported by data from the knowledge base One filter can always be bypassed. That is why we do not rely on just one layer. Injection filter blocks manipulation before the model call, RAG prevents the model from inventing without data, and semantic filter checks that the answer does not go too far from the knowledge base. Each level closes the blind spots of the previous one. Together, they work as a system. 𝗢𝗻 𝗙𝗲𝗯𝗿𝘂𝗮𝗿𝘆 𝟭𝟰 𝗧𝗿𝘆𝗹𝗹 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 𝘄𝗶𝗹𝗹 𝗯𝗲 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝗼𝗻 𝗦𝘁𝗲𝗮𝗺, 𝗮𝗱𝗱 𝗶𝘁 𝘁𝗼 𝘆𝗼𝘂𝗿 𝘄𝗶𝘀𝗵𝗹𝗶𝘀𝘁 - https://lnkd.in/d6HsnAZT

We're proud to share that we've received an Epic MegaGrant from Unreal Engine Under the leadership of our CTO, Vladimir Belyaev, we are working to bring on-device AI into Unreal Engine games! Thanks to the support from the MegaGrant, we are officially one step closer to delivering this product to our clients. If you want to be among the first to integrate seamless on-device AI into your projects, head over to our website and join the waitlist: https://tryllengine.com/ #EpicMegaGrants #UnrealEngine #UE5 #UEFN

𝗧𝗿𝘆𝗹𝗹 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 is the first fully on-device AI gaming assistant - 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝗳𝗼𝗿 𝗳𝗿𝗲𝗲 on Steam starting 𝗙𝗲𝗯𝗿𝘂𝗮𝗿𝘆 𝟭𝟰𝘁𝗵. No more alt-tabbing or disconnecting to look things up. 𝗔𝘀𝗸 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀 and get answers powered by community- and developer-approved information - 𝗮𝗹𝗹 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗹𝗼𝗰𝗮𝗹𝗹𝘆 on your device. We're starting with a community preview for Project Zomboid and Stardew Valley. The goal is to work closely with players, gather feedback, and refine the experience before expanding to more titles. For us at Tryll Engine, this is also a way to stress-test our 𝗼𝗻-𝗱𝗲𝘃𝗶𝗰𝗲 𝗔𝗜 stack across real hardware and real player sessions - the kind of data you can't replicate in an office. Which games would you like to see supported next? Let us know in the comments! Wishlist on Steam: https://lnkd.in/dQvipHYy