ZK Powered AI Agents: Revolutionizing DeFi Yield Strategies | Andrei David - Summerfi

[Music] um so hello everyone um I hope you're having a fantastic time here at e Denver and uh it's truly incredible to see so many Builders and innovators gather here uh I'm Andre city of um sumeri uh and today I'm going to share how we are working to revolutionize uh yield optimization in defa using AI agents and uh ZK proofs uh this is a quick 20 minutes talk uh but uh if anything I share picks your interest uh please come chat with me afterwards I'll be around and I'd love to hear your thoughts or explore potential um uh collaborations uh and now just a quick disclaimer um I know you've probably seen a slide like this one 100 times here at Denver but uh I need to go through this uh this presentation is not Financial or legal advice uh we are being building a complex and novel system here and I want to emphasize that some of what I'll show today it's still working progress and while we are moving quickly with development uh we are taking a measured approach to production deployment um our philosophy is to roll out these features uh gradually as our testing and auditing concludes uh and we are confident in the security and um reliability of um all this um uh in inovative Innovative functionalities so now let's dive into the exciting stuff um but before going into the technical details let me briefly introduce summery we are a platform focused on automated yield optimization across the most trusted protocols in defi we are currently operating on ethereum mainnet based and arbitrum and we've just closed crossed $40 million in um tvl in the in Less Than 3 weeks since uh since launch uh what sets us apart is our unique approach to how we uh develop ELD strategies and how we execute them and I'll talk in more detail about that uh during the during the presentation um so um I won't have time to discuss in detail the vaults and risk manager uh pillars of our architecture um but I think it's important to at least give you a high level of how they interact with each other uh to keep the system in Balance um so our architecture is built uh on three pillars each with a very specific role and responsibility uh and this separation of concerns uh isn't just architectural Elegance it's practical security where each component is an export in its domain um so I'll briefly explain each pillar to give you a better understanding of how they work um together and the first pillar is the vaults uh our vaults follow the ERC 4626 uh standard and they are governed by the summer token and its community and what's crucial here is the separation of powers summer holders vote in on which protocols each vot supports uh but they don't set the risk the risk parameters uh this ensures um governance focuses on strategy and expansion without compromising uh risk management uh the second pillar is the risk risk manager and for us block analytica serves as our independent risk Guardian they are the ones building the risk models for each of those um those vaults uh they set the deposit caps for the vaults maximum Market exposure rebalancing flow controls and so on and this independent risk management layer ensures that uh no matter how aggressively our yield seeking strategies are they always operate within the prudent risk boundaries set by block analytica and uh finally the third pillar the AI keeper Network um unlike traditional yield aggregators using manual or multi- uh rebalancing uh our AI Keepers react almost instantly to Market changes operate continuously 247 have one uh execute within strict risk parameters set by by the risk curators and compete to find optimal optimal strategies and we'll discuss a little bit more about how they compete for those strategies um uh next um so this completes the uh this complete system represents our vision for the future of um defy yield optimization automated decentralized secure and constantly improving through competitive strategy development um and now before moving forward let me step back and talk from first principles about the problem that we are trying to solve uh a problem that is C to all economic systems is not uh particular to defy or threat F and um the definition that we have here on the slide basically States this idea that like economic systems are price Discovery mechanisms that optimize the allocation of capital across different risk reor profiles and with summery this is exactly the problem that we are trying to solve for defy where to allocate Capital to extract the maximum value out of um uh rates for underlying underlying protocols um and um just to kind of give you a a bit of a sense of the scale of the problem that we are trying to to solve in defy this is a recent tweet from from defile Lama about stable coin yields alone and uh they have their 1,800 pools 460 protocols on 101 chains right and this is just stable coins um so this level of fragmentation makes it practically impossible for any human to manually track all these opportunities and manually allocate Capital uh across this risk reward curve that we are talking about so let's talk about how we are solving um for this uh uh at summery by using um AI powered agents and when we talk about AI in this Con in this context we have a very specific definition uh AI is the ability of a system to continuously analyze complex data patterns derive actionable insights and autonomously execute optimization decisions so this is not General AI this is not an AI chatbot it specialized intelligence designed specifically for yeld optimization uh problems uh and we believe defi yield optimization can only be solved at scale effectively with these capabilities that AI brings to the to the table um and um just to give you a brief overview of some of the subsystems that are part of this AI agents um I want to share some of our work in eeld forecasting here uh with the developed machine learning uh algorithms that analyze historical yield data and predict future movements across across various underlying protocols uh and this chart shows our forecasting models prediction compared to actual yields for several major lending protocols you have AER fluid compound and um and so on um and in our testing we've seen amazing results with prediction accuracy within 15 to 20 basis points for up to hour long forecasts um and this forecasting capability allows us to be proactive rather than reactive to the market instead of responding to yield changes after they occur uh our AI agents will anticipate movement and position funds optimally before major shifts happen um another critical AI application we developed is um anomaly detection for price and uh rate feeds uh and as you probably know defi is vulnerable to various forms of Market manipulation and Flash loan attacks uh dep pegging of stable coins that can temporarily distort prices and yields and basically in real time change the risk profile of specific Capital allocations uh our proof of concept anomal detection system monit s feeds in real time comparing incoming data against expecting expected ranges and historical patterns and when it spots outliers or anomalies it can flag them for review or suggest posing operations and all of this is is happening in in in real time and it's one of the again the core components of the of the AI agents um and this is another I think interesting one uh multiple AI validator agents independently analyze the proposed action and try to reach consensus uh you have here an example output from um our system operating on arbitrum uh just a couple of days back so this is this is production data that the AI agents are responding to in in in real time and you can see there in um uh natural language the analysis of all the buffer management implications for that rebalancing operation API optimizations and the assessment uh includes concentration risk um protocol compliance with deposit caps flow limits and all the risk parameters that I was telling you about earlier um and in this case both agents approved the operation based on rigorous analysis of benefits versus risks um and this multi-agent approach creates additional security through consensus and provides detailed documentation of of decisions that are made by those AI agents um and um this I think gives us the best of both worlds uh the speed and efficiency of automation with the prudence and transparency typically associated with um human human governance um the output being in natural language it can be easily understood by um either human operators um monitors or um or customers so so we've seen a little bit about how AI can uh revolutionize yield optimization through continuous analysis pattern recognition and autonomous execution um but um how can we improve of this even further um how do we coordinate and reach consensus on choosing the best strategies and um execute them and the biggest question becomes how do we enable decentralization of this process to attract the best agents AI or um human plus AI uh and reward them um without compromising any of the security guarantees of the protocol um and in this case decentralization isn't just a buzzword for us uh it's core to our approach and the AI agent component of the protocol is the last one that's still centralized right now but but we are working hard to to to decentralize it as well and we believe that most robust yield optimizations will come from harnessing collective intelligence of many participants in the ecosystem rather than relying on a single centralized decision maker so we are creating this decentralized marketplace where the best yield strategies can be created for the benefit of sumeri users bringing together the best intelligence and best strategies in service of optimal optimal returns um so our meta market concept is an evolution of field aggregation from manual rebalancing to autonomous Marketplace of competing strategies the agent aren't just passively monitoring they are actively competing to find the most efficient Capital allocation path across the defi ecosystem and then rewarded for that uh but there's a fundamental challenge here in traditional markets um strategies are valuable intellectual property and we believe this holds through in defy as well um if we force agents to reveal their strategies before selection we create two problems front running and strategy theft and these kinds of incentives um minimize the appetite for any AG agent or any uh player to to go and invest in coming up with those strategies so what we are doing and I think you can think of it like a sealed bid action for for strategies where agents submit a proof that certifies two things validity so the fact that the strategy meets all the protocol rules and risk parameters and performance uh how it improves overall yield Beyond a specific uh threshold um and the key is proof the proof the ZK proof proves the these facts without revealing any strategy details until after selection and after the agent is rewarded by the system uh strategies that are not selected for execution continue to remain private uh preventing strategy theft and diminishing uh competitiveness in future um uh auction rounds um we've already deployed our first proof of concept Noir circuit and solidity uh verifier on base um and you can see here a little bit more details about uh the inputs to the Noir circuit uh those are public inputs like um deposit caps interest flows risk parameters and they are visible to all the to all the agents uh and the Agents privately created their Moves In in a very simple format Source destination protocol and the transfer amount so basically solving for this like core problem of reallocating assets to maximize for uh for yield uh the Noir circuit verifies risk constraint compliance flow limits and also performance performance outputs for those uh rebalancing operations um and this approach delivers three critical advantages um complete front running protection through strategy privacy uh and through incentives for strategy Innovation um and now looking ahead uh our road map focuses on several key priorities first we are expanding our AI capabilities second we are actively working on decentralizing the agent pillar as I've mentioned and this is a critical step towards our vision of what truly decentralized market for ELD strategies where independent agents can compute and be rewarded uh and we are doing this through ZK through ZK proofs um so just to quickly summarize what I've presented today uh we are creating a paradigm shift in how yeld optimization Works in defi from siloed manual processes to an efficient meta market where agents compete to deliver uh optimal Capital allocation um our vision for the agents pillar is to create the first through meta market for yield strategies uh where agents compete not just on performance but on Innovation as well uh think of it as an evolution of yeld aggregation from manual rebalancing to autonomous Market Place of competing strategies and while we are still in development we are already seeing how this meta market concept could revolutionize uh defile deal optimization uh at an ecosystem level the combination of AI for strategy Discovery and ZK proofs for trustless verification creates a powerful new model for defi uh by solving the critical problem of strategy privacy and provable performance we are enabling a true open ecosystem where the best ideas can win on Merit alone regardless of who creates them uh and this is this isn't just about incremental Improvement to already existing models it's about fundamentally rethinking how yield optimization can work in a centralized World in a decentralized world um and ultimately what we are building here goes beyond just yeld optimization um um remember how we started uh with that uh definition about economic systems that are price Discovery mechanisms uh that optimize the allocation of capital across different risk rewards profiles um today we are solving for yield optimization in defi uh but tomorrow the same approach can revolutionize Capital allocation across all markets from def5 to trillion doll um traditional Finance sectors uh we are creating the infrastructure for more efficient markets everywhere Capital needs to find its optimal risk reward balance um thank you for um your attention and please um join our community this was a super short talk given the the um the the wise subject from AI to ZK proves um but I'll be around if you have any questions or you want to discuss potential collaborations um just come back and I'll be happy to I'll be happy to take question and um dive deeper into the details of the of the implementation thank you [Applause]