In this episode of "Exploring the AI Rainforest," host Derek E. Silva interviews Himanshu Udupi, a sophomore at the University of Illinois at Urbana-Champaign (UIUC). Himanshu is making significant contributions to Zeru's zScore research, working on merging onchain and offchain reputation systems into a portable value metric. His fresh perspectives on the intersection of AI and blockchain technology are helping shape the evolving landscape of decentralized technologies.
Born and raised in Bangalore, India, Himanshu completed his schooling there before deciding to pursue higher education in the United States. Following his mother's advice to take the SATs, he performed well and was accepted to UIUC, his top choice due to its prestigious research reputation. Himanshu appreciates the research-focused and hustle-driven atmosphere at UIUC, which motivates him to contribute meaningfully through his academic pursuits.
His introduction to computers began in elementary school with animation software called "pivot stick figure." By sixth grade, he had his own laptop and built a website for his mother's birthday, which sparked his interest in computer science. Himanshu mentioned that Iron Man's AI assistant Jarvis was a significant inspiration for his interest in artificial intelligence.
Himanshu's journey with Zeru began during the summer after his freshman year. At his mother's suggestion to be productive during his break, he interviewed at a startup incubator lab near his home in Bangalore. During the interview, he was offered several project options and chose one related to identifying fraudulent blockchain transactions. On his first day, he met Guru and Akshay, the founders of Zeru, and they immediately connected. What started as a summer internship evolved into a full-time role.
The initial version of the zScore model focused on analyzing Aave data, a lending protocol across various blockchain networks. Himanshu described his process of transforming approximately 10 million transaction records into meaningful patterns:
Data Exploration: He spent a month understanding the data structure, converting transaction-level data to wallet-level representations, and comprehending repayment mechanisms.
Pattern Recognition: Despite initial concerns about data spread, Himanshu discovered clear trends and patterns in user behavior, with most data centered around averages and minimal deviation.
Clustering Approach: Rather than assigning individual scores to each wallet, the team opted for a clustering model that groups similar behaviors together and assigns score ranges to each cluster.
Scoring Methodology: For each cluster, they mapped minimum feature values to the minimum score range, maximum values to the maximum score, and developed a system to score users between these extremes based on the most important features for that particular cluster.
The team is currently working on zScore v2, which will incorporate data from additional sources:
Other lending protocols beyond Aave
Decentralized exchange (DEX) data
DAO voting information
This expansion presents significant challenges in consolidating different types of blockchain behavior into a unified scoring system. While the basic framework remains the same, the algorithms and inferences drawn from clustering need to be reimagined. Himanshu emphasized that wallets engaging in extractive behaviors that benefit themselves at the community's expense will be penalized in the scoring system, as the model values contributions to the wider ecosystem.
Himanshu spoke highly of his interactions with Zeru's advisors, particularly Piet Martens and Parag Paul. He maintains an 18-email chain with Piet, who provides timely and constructive feedback on his papers. Parag has helped Himanshu develop not only theoretically but also practically, with guidance on presenting at conferences and evolving as a researcher. Both advisors have been accessible and supportive, regardless of how basic Himanshu's questions might be.
Looking ahead to the next 12 months, Himanshu is most excited about the release of zScore v2 and seeing how it performs across different protocols. He sees potential for their work to contribute significantly to both Web3 and AI fields, as they're working with cutting-edge research published within the last few months. Additionally, he looks forward to the potential launch of the Zeru token, though he acknowledges this decision is not in his hands.
Himanshu's work with Zeru represents an important step toward creating a more comprehensive and equitable reputation system for the blockchain world. By incorporating both onchain and offchain credentials through platforms like Reclaim Protocol, the zScore model could potentially become a more effective credit scoring system than traditional methods, especially in regions where such systems don't exist or are flawed. As this technology continues to evolve, Himanshu remains at the forefront, helping to shape the future of decentralized finance and AI integration.
Listen or subscribe now on Apple Podcasts, Spotify, or your favorite podcatcher.