Kwindla Hultman Kramer is the co-founder and chief executive of Daily, the real-time communications company behind Pipecat, an open-source framework for conversational voice and multimodal applications. His career spans distributed software agents, spatial computing, WebRTC infrastructure, and the challenge of making AI conversations fast enough to feel natural.
At the MIT Media Lab, Kramer worked on networked devices and programmable systems and co-authored research on decentralized agents for network routing with Nelson Minar and Pattie Maes. He subsequently helped build AllAfrica’s web infrastructure, worked on technology projects for the Democratic National Committee and Media Matters for America, and advocated open-source software for public institutions.
In 2006, Kramer co-founded Oblong Industries with John Underkoffler to commercialize spatial and gestural computing associated with Minority Report. Its work on spatial, multi-user, multi-screen computing explored how people, devices, and shared visual environments could operate together.
As conversational models improved, Daily extended that communications stack into Pipecat, an open-source, vendor-neutral framework that combines live audio, transcription, model inference, speech synthesis, and external tools. Hosted infrastructure followed as Pipecat Cloud, designed for deploying and scaling production voice agents.
Voice-to-voice latency defines conversational quality. Kramer measures the interval between someone finishing a sentence and hearing an agent respond. Human exchanges often turn within roughly 500 milliseconds; production systems may target approximately 800 milliseconds across audio processing, networking, transcription, inference, and speech generation. His low-latency architecture emphasizes measuring every stage and colocating dependent services.
Open infrastructure preserves architectural choice. Pipecat’s vendor-neutral voice-agent orchestration lets developers swap models, speech services, transports, and telephony providers. Pipecat Cloud addresses cold starts, autoscaling, regional deployment, observability, and long-running sessions; Kramer also supports direct WebRTC connections when self-hosting makes sense.
Useful agents must listen, interrupt, and act. Reliable systems need turn detection, interruption handling, background-noise resilience, conversational context, and asynchronous tool calls. Kramer has emphasized patient voice interfaces for healthcare, interviews, and software onboarding, including conversational healthcare intake that can replace cumbersome forms.
Audio-native models must earn their place. Direct speech-to-speech systems can preserve vocal and multilingual information lost in transcription, but text-centered pipelines may still provide stronger instruction following and tool use. Kramer treats that choice as an application-specific production reliability tradeoff.
Agents point toward AI-native software. His open-source multiplayer experiment Gradient Bang explores asynchronous subagent orchestration, persistent context, generated interfaces, and conversational control: software coordinating fast human interaction with longer-running background work.
An interruptible joke bot, a patient-intake call, and a geography game reveal how media transport, conversational state, model placement, and measurement shape voice AI latency.
A responsive voice agent needs more than a good model: it needs programmable media handling, deployment suited to live conversations, and careful choices about where inference runs.
From Bush’s Memex to a multiplayer game built around LLMs, the path beyond agents depends on new ways to combine computation, context, interfaces, and human intent.
How do you imagine software before its infrastructure exists?
From conference Wi-Fi to a talking Raspberry Pi, natural voice interaction depends on a tight latency budget, the right media transport, and room to experiment.
▶ Watch ↗
