mitramandal.ai EP4 - Pune

​We're hosting the fourth mitramandal meetup. same vibe: builders, researchers, and tinkerers sharing real-world AI workflows.

​Speaker lineup:

​1. Adaptive Load Balancing for LLMs

Akshay Deo, Co-Founder & CTO @Bifrost & Maxim AI

​LLM providers rarely fail cleanly, they just get slow or start rate limiting you. This talk shows how an adaptive load balancer in Bifrost watches each provider's real behavior and steers traffic away from the struggling one before users notice.

​2. ​Running Untrusted Agent Code at Scale
Everyone's building agents, but how do you deploy them securely and keep the data with you?

Shekhar Patil, AI, Infrastructure and Partnerships @Last9

​Your AI agent runs code nobody on your team wrote: generated by an LLM, pulled from a registry, or shaped by a prompt-injected webhook. The moment it can call tools and spend money, you're running a multi-tenant code-execution platform with an adversary inside it. This talk shows how we deploy untrusted agents safely by making isolation structural, not conventional: enforced by the kernel, the crypto, and the database key, never by trusting the agent to behave. You'll leave with a concrete threat model, copyable controls (kernel sandboxing, per-agent crypto, pre-call budgets, tamper-evident audit), and the one test that decides them all, does a control make the attack impossible, or just tedious? It's zero trust for agents, and we'll show where a separate-kernel micro-VM finishes the job.

​

​3. Architecting Security-First Multi-Agent Systems: Enforcing RBAC, PBAC, and Data Sovereignty

Vivek Singh, Sr Technical Leader @Cisco

In multi-agent architectures, security and compliance cannot be treated as an afterthought or an external layer. For enterprise-grade systems, dynamic data access boundaries must be fundamentally woven into how agents discover and interact with tools and data sources. This session provides an architectural deep dive into how we enforce Role-Based Access Control (RBAC) and Permission-Based Access Control (PBAC) directly within multi-agent orchestration layers. We will explore how to leverage the Model Context Protocol (MCP) to build a governed ecosystem where agents explicitly inherit user-level and context-level permissions. Crucially, we will address a core architectural challenge: Data Sovereignty vs. Insights Synthesis. You will learn how to design systems where raw data must strictly reside within specific regional boundaries (e.g., country-specific compliance laws), yet agents can still safely cross-compile and synthesize high-level insights for global orchestration. We will map out the architectural blueprints, critical edge cases to consider, and the open engineering challenges in building fully compliant, security-aware AI systems.