Ankush Rastogi

Ankush Rastogi is a senior data solutions engineer at Prosodica who builds AI systems for analyzing customer conversations, evaluating contact-center employees fairly, and managing large catalogs of agent tools. His work centers on a practical architectural question: how to give models the right evidence and capabilities without sacrificing accuracy, privacy, or accountability.

In 2024, Rastogi worked as a senior business analytics engineer at Prosodica, helping apply generative AI to contact-center quality assurance as the company expanded beyond smaller transformer-based speech and voice models. His account of that transition describes moving from selective manual call reviews toward broader automated analysis.

A 2025 white paper on conversational intelligence developed his approach further: reliable telephony analytics requires coordinated audio ingestion, transcription, model integration, security controls, and human oversight. By 2026, he led the data and evaluation architecture behind Prosodica’s Auto-Evaluation, which assembles behavioral signals, customer-effort indicators, and conversation outcomes into structured evidence before a language model generates an assessment. The design makes automated employee evaluations easier for human reviewers to understand and challenge.

• Semantic routing for scalable agents. With Prosodica colleague Sohail Shaikh, Rastogi examined why supplying every tool schema with every request increases token costs, slows responses, and complicates tool selection. Their AI Engineer World’s Fair session outlines an alternative: index tool descriptions, retrieve a small relevant subset, and evaluate accuracy, latency, cost, and retrieval failures together.

• Just-in-time context injection. Rastogi applies lazy-loading principles to agent infrastructure: determine the user’s intent first, then inject only the necessary tool definitions. He emphasizes practical thresholds and monitoring; small catalogs may not need routing, while larger systems should test different retrieval sizes against real workloads.

• Developer-side privacy. His PromptShield project is a Visual Studio Code extension designed to detect and sanitize sensitive information before it reaches AI assistants.

• Local-first incident analysis. His SignalGraph project maps relationships among operational events, helping teams investigate cascading incidents and downstream effects while keeping their data local.