Anjaneya Sastry Kappagantu
Anjaneya Kappagantu is an Expert Solution Architect at AT&T with 25+ years architecting enterprise and telecom (BSS) systems — including as AT&T's solution architect for the first-ever iPhone launch and end-to-end Apple integration. He mentors principal architects and advises leadership on GenAI and agentic AI.
Alongside his career, he is an AI educator and researcher focused on local and edge AI — fine-tuning models on Apple Silicon with MLX, designing agentic AI workflows, and researching test-time adaptation, including a human-in-the-loop method he developed, Reinforce-or-Retrain (RoR). He mentors students and product teams, speaks at AI and engineering conferences and meetups, and has judged 150+ AI pitches. He also explores non-standard architectures such as Kolmogorov–Arnold Networks (KANs) applied to Pāṇinian Sanskrit grammar. He is a trained aerospace engineer and a professional tabla player.
LinkedIn: https://www.linkedin.com/in/anjsastry
Autonomous agents are the newest — and least predictable — non-human identities in the enterprise: NHIs that compose their own commands. Scoped credentials shrink their blast radius, but even attribute-rich IAM answers whether an identity may act — never whether what’s inside can ever come back. This session presents an open-source pre-tool-call hook that closes the gap: a deterministic blast-radius floor plus a fully local “simulated-grief” grader, merged on a five-rung ladder where the stricter answer always wins. In testing it caught customer data that permissions allowed — even hidden under renamed directories — and the architecture independently re-derives a millennia-old Sanskrit analysis of attachment: code first, verses after. It ships today, open source: fewer approval prompts where they’re noise, harder stops where they’re load-bearing.
Key takeaways: