How Hari Kishan Defends AI Architecture Over Bigger Models for Better Voice AI

Episode Summary:

In this episode of Engineering Choices You Have to Defend, host Nicola Onassis sits down with Hari Kishan, an enterprise technology and AI leader, to discuss why modern AI systems require thoughtful architecture, specialized models, and strong human oversight rather than simply relying on larger language models.

Hari shares his journey from business intelligence development to Python, startup leadership, enterprise AI, and large-scale technology transformation. He explains how modernizing legacy contact centers with dynamic routing, conversational AI, voice biometrics, sentiment analysis, and automated call summaries can improve customer experience while reducing unnecessary friction.

The conversation explores why latency is one of the most critical challenges in voice AI. Hari explains how even small delays can affect customer trust and why newer speech-to-speech models can deliver faster, more natural interactions. He also discusses the security risks of voice biometrics and the importance of combining automation with strong cybersecurity safeguards.

Hari explains an AI architecture that combines small language models with specialized context lakes, confidence scoring, and large language model fallbacks. Rather than asking one model to handle every task, this approach gives specialized systems defined responsibilities while improving accuracy, relevance, and reliability.

The discussion also examines the role of humans in AI-powered systems. Hari emphasizes that AI should enhance employees rather than remove human oversight, with people remaining responsible for validating important outputs and enforcing governance and security controls.

For engineering leaders building AI systems in production, this episode offers practical lessons on voice AI, latency, specialized models, agentic architecture, security, governance, and why getting the architecture right matters more than simply choosing the latest model.

Key Takeaways:

• AI modernization can dramatically improve the customer experience of legacy contact centers.
• Dynamic routing can reduce frustrating IVR experiences and connect customers with the right agents faster.
• Voice biometrics can streamline authentication but must be protected with strong cybersecurity controls.
• Latency is critical in voice AI because even small delays can damage customer trust.
• Speech-to-speech models can reduce the complexity and latency of traditional voice AI pipelines.
• Small language models can provide more focused and domain-specific responses.
• Context lakes can provide AI agents with relevant organizational knowledge and improve over time.
• Confidence scoring and fallback systems can improve AI accuracy and reliability.
• Specialized AI agents can perform specific tasks more effectively than one model trying to do everything.
• AI should enhance human workers while keeping people responsible for validating important outputs.
• Sensitive data such as PII should be protected through strong architectural controls.
• Architecture should come before model selection, with models treated as components within a larger system.

Connect with Hari Kishan:

LinkedIn: linkedin.com/in/hari-kishan

Website: thethinkingvoice.com

Listen Now & Subscribe:

Apple Podcasts, Spotify, Amazon Music, YouTube, iHeartRadio, Captivate, or wherever you get your podcasts.

Engineering Choices You Have to Defend explores the real technical decisions behind AI systems, enterprise architecture, and scalable software engineering.