Akila Balasubramanian

Semantic Condensation of High-Cardinality Time Series for LLM-Driven Observability

Abstract:

Modern cloud observability platforms generate high-cardinality time series data that is inherently difficult to interpret at scale and poorly aligned with the reasoning constraints of large language models (LLMs). Raw telemetry, often comprising thousands of time series (TSIDs) with dense temporal sampling, leads to token inefficiencies, obscured signals, and contradictory narratives when consumed directly by LLM-based assistants. 

This paper introduces Semantic Condensation, a token-aware transformation layer that converts large-scale time series data into structured, semantically consistent digests optimized for LLM-driven troubleshooting. The approach combines vectorized statistical pre-analysis, multi-signal importance scoring, adaptive token-budget-aware tiering, and behavior-aware trend classification to preserve critical operational signals such as anomalies and change points while minimizing representation cost. 

Unlike traditional downsampling and aggregation techniques, Semantic Condensation explicitly optimizes for LLM interpretability and narrative coherence, ensuring that derived summaries remain internally consistent. Experimental evaluation demonstrates that the approach scales to 10,000+ time series within strict latency constraints, reduces token footprint by orders of magnitude, and significantly improves downstream LLM reasoning accuracy. This work establishes a new abstraction layer for observability systems: LLM-aligned semantic representations of telemetry data.

Profile:

Akila Balasubramanian is a Software Engineering Technical Leader at Cisco, where she leads the design and development of AI-powered observability and intelligent troubleshooting capabilities for cloud-native platforms. Her work spans distributed systems, agentic AI, observability, and platform engineering, with a focus on building production-ready AI systems that help engineering teams detect, investigate, and resolve complex operational issues more effectively.
Over the years, Akila has led the development of technologies including AI-directed troubleshooting, automated root cause analysis, real user monitoring, session replay, telemetry intelligence, and intelligent investigation workflows. Her work emphasizes trustworthy AI, explainable reasoning, robust evaluation, and scalable platform architecture that can support enterprise-scale production environments.

Akila is passionate about bridging research and real-world engineering, translating advances in AI into reliable systems that solve meaningful customer problems. She frequently shares practical lessons on building agentic AI platforms, LLM evaluation, observability, and the engineering challenges of deploying intelligent systems at scale.