Ankit Anand

Bounded-Space Inference: Overcoming the LLM Memory Wall via Neuro-Symbolic Architecture

Abstract:

As generative artificial intelligence systems scale to accommodate massive context windows, the Key-Value (KV) cache of the Transformer architecture exhibits an unsustainable linear memory scaling constraint. This bottleneck rapidly exhausts High-Bandwidth Memory (HBM) in modern computing environments. This keynote explores a hardware-software co-design paradigm that achieves bounded-space memory scaling. By leveraging Vector Symbolic Architectures (VSA) and physically routing probabilistic tokens separately from structurally rigid tokens, systems can maintain high-fidelity logic retention without the prohibitive memory overhead of traditional KV caching. The session will detail the mechanics of attention-gated compression and leaky bundling, illustrating a viable pathway for deterministic integrity in autonomous AI-agentic enterprise operations.

Profile:

Ankit Anand is a Managing Consultant and Data Management Architect leading the Data Quality practice at Syniti. With extensive experience architecting enterprise data systems across complex regulated environments, his research focuses on neuro-symbolic architectures, deterministic AI governance, and advanced AI infrastructure. He is the inventor of the Neural-Holographic Cache Controller and holds four pending utility patent applications with the USPTO related to AI-agentic operations and hardware caching. He is the author of the forthcoming book "The Deployed Data Scientist" with Technics Publication and is a 2026 Global Recognition Award recipient for exceptional achievement in research. He holds a Master of Science in Business Administration and a Bachelor of Engineering in Computer Science. He is a Senior Member of IEEE and a Professional Member of ACM.