← All speakers

Bio, Work & Ideas

Nagkumar Arkalgud

Conference affiliation: Microsoft · 2025

Nagkumar Arkalgud is a senior software engineer at Microsoft working on Foundry Observability, the infrastructure that helps developers evaluate, secure, and monitor AI applications. His projects bring adversarial testing and standardized tracing into everyday engineering workflows, including systems spanning multiple clouds.

Earlier, as a Capital One software engineer, Arkalgud helped create Odis, a voice-controlled music-production assistant built at TechCrunch Disrupt 2017. The Alexa-to-MIDI system combined an Alexa skill, AWS Lambda, WebSockets, and a Node.js client so musicians could operate recording software hands-free; the project won an Amazon Alexa hackathon award.

At Microsoft, Arkalgud helped build the Azure AI Evaluation SDK, making capabilities from the open-source PyRIT security framework accessible through programmable scans and hosted results. His AI Engineer World’s Fair demonstration showed developers how to test a PostgreSQL-backed retrieval application, locally hosted Ollama models, and Azure OpenAI deployments against adversarial prompts.

By Microsoft Build 2026, he was co-owning projects for cross-cloud agent observability and open-source Foundry integration.

  • Practical adversarial evaluation: Configurable risk categories and composable transformations, including encoded or reversed prompts, let developers systematically expose failures and inspect successful attacks.
  • Application-level and model-level testing: Separate scans of retrieval applications and underlying models help teams identify weaknesses and assess how different guardrail configurations change outcomes.
  • Cross-cloud agent observability: OpenTelemetry GenAI conventions and W3C trace context connect agents deployed across Azure, AWS, and Google Cloud, linking routing, latency, and evaluation results.
  • Open-protocol interoperability: Foundry integrations combine LangGraph, Model Context Protocol tools, reusable skills, browser automation, and standardized telemetry without abandoning familiar open-source components.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

A local RAG application becomes a target for interactive attacks and automated scans, showing how to inspect failures and connect evaluation results to guardrails.

  • What happens when users push past the intended behavior?
    1:10 ↗
  • Putting specialist red teaming tools within reach
    4:03 ↗
  • An interactive attack loop against a local RAG application
    5:36 ↗
  • Turning the interaction into an automated scan
    8:45 ↗
  • Inspecting application failures in the dashboard
    11:03 ↗
  • Scanning a model directly
    11:56 ↗
  • Connecting evaluation to deployment decisions
    14:22 ↗
  • Where the guardrails sit
    17:14 ↗

References