Krishna Suman Dara

Title of the Talk: Why Small Language Models Belong at the Network Edge

Abstract :
Large language models dominate the AI conversation, but production networking and AIOps systems rarely need — or can afford — the capability of a frontier model. This talk makes the case that Small Language Models (SLMs) are not a compromise but the correct architectural choice for the environments network engineers actually operate in: routers, edge devices, and real-time telemetry pipelines where latency, privacy, and compute budgets are non-negotiable constraints, not nice-to-haves.

Drawing on recent advances in model compression — quantization, pruning, and knowledge distillation (including Minitron-style teacher-correction approaches that cut training cost by two orders of magnitude) — this talk breaks down how SLMs like Phi-4-mini and Gemma now deliver near-frontier reasoning at a fraction of the footprint, enabling on-device inference without a round-trip to the cloud. It then turns to why this matters specifically for agentic AIOps: every autonomous Thought→Action→Observation loop in an agent-driven network operations pipeline compounds latency and cost with each iteration, making model size a multiplier on operational risk, not just an infrastructure line item.

The talk closes with a practical framework for deciding when a narrow, fast, privacy-preserving SLM beats a general-purpose LLM in network automation — and when it doesn’t — grounded in real trade-offs around instruction-following limits, domain specialization, and the compliance demands of telemetry data that should never leave the edge.

Bio :
Suman Dara is a Senior Principal Engineer at Hewlett Packard Enterprise (HPE) / Juniper Networks, where he leads R&D initiatives in advanced routing architecture and product engineering. With over 20 years of experience in network architecture, distributed systems, and routing R&D, Suman has been instrumental in the development and launch of flagship routing portfolios, including the ACX and MX product lines.

Suman bridges deep infrastructure expertise with emerging AI/ML methodologies to solve complex, non-linear network challenges. He is an active industry innovator, focused on patenting and deploying Network Digital Twin technologies to enable autonomous remediation and predictive network analytics.