simalexan

Journey to Intelligent Conversational Agents

A talk Konstantinos Siaterlis and I gave at AWS Community Day Adria 2026 in Ljubljana, Slovenia, on October 22, 2026: why AI has us working more rather than less, what a voice agent is made of, and how we built Ultron, a multimodal voice agent that listens to a design discussion and draws on a board.

Click the slides, then use ← and → to move and F for full screen. Open the slides on their own page →

What it covers

  • Opening: why we now work more, not less, with AI (the Jevons paradox, and “botsitting” ten agents), and whether voice agents are a cure.
  • Strands: building voice agents with Strands and Amazon Nova Sonic, from ten lines of code to an agent that researches a topic and emails it to you.
  • Origins: from Turing to ChatGPT, by way of the re:Invent 2018 workshop “Alexa, ask Jarvis to create a serverless app for me”.
  • Anatomy: what a multimodal voice agent is made of, and why we chose each part of Ultron.
  • Close: Ultron drawing its own architecture, Wardley Maps, and three key points.

Three key points

  • Voice in, pictures out: speak to tell it, see to understand it, both ways in one conversation.
  • Built from parts: diarization, interruptions, speech, a voice and memory, each chosen on evidence.
  • Fewer calls, more decisions: an agent that calls you only to decide cuts the waiting and the checking.