← All speakers

Kwindla Hultman Kramer

Conference affiliation: Daily · 2026

Kwindla Hultman Kramer is head of open source at Deepgram and the co-founder and former CEO of Daily, where his team built Pipecat. The framework gives developers a way to build voice agents from interchangeable speech and language services while handling the ordinary complications of conversation, like a caller pausing to think or cutting in to correct an answer.

On this page

From programmable tiles to shared screens

In 1995, Kwin wrote web pages by hand, built servers and search software in C, and made publishing tools in Perl. At the MIT Media Lab, he explored what could happen when software itself moved between devices. For his 1998 master’s thesis, supervised by Mitchel Resnick, he studied how small programmable objects called Tiles could pass running programs to one another. With Nelson Minar and Pattie Maes, he also studied how cooperating software agents could adapt routes through a network.

As founding CTO of AllAfrica, Kwin helped build the infrastructure that brought reporting from many African newsrooms into one publishing service. His civic technology work included projects for the Democratic National Committee and Media Matters for America. He advocated for open-source software in public institutions so they could spend more on training and less on licenses. AllAfrica · Kwin’s career account

In 2006, Kwin and John Underkoffler co-founded Oblong Industries to turn spatial-interface research into a product. Underkoffler’s MIT work had supplied the basis for the gestural interface he helped create for Minority Report. Kwin ran Oblong as CEO while Underkoffler served as chief scientist. Their g-speak Spatial Operating Environment let people use gestures and work together across screens and devices. In 2026, Kwin still described building a shared, multimodal computer as work he was pursuing. Kwin and Underkoffler on g-speak · Kwin’s 2026 account of AI-native software

How Daily’s video infrastructure became a voice framework

Kwin and Nina Kuruvilla co-founded Daily and took it through Y Combinator’s Winter 2016 batch. They expected to sell their video conferencing appliance in the United States first, and international orders arrived on the first day. In March 2019, Nina introduced Daily’s API for embedding video calls in other products. Daily’s global media network let callers connect to nearby servers and carried streams between those servers, so people on different continents could share a call without all connecting to the same distant location. Daily announced a $40 million Series B in 2021, and Norwegian telehealth company Confrere joined it in 2022. Early international orders · Nina’s video API launch · Daily’s global mesh · Series B · Confrere

After GPT-4’s release in 2023, Kwin and his Daily colleagues used WebRTC to connect speech recognition, a language model and synthesized speech. The same problems kept surfacing. A pause could mean the caller was thinking or ready for an answer. If the caller interrupted, the software had to keep track of how much of the reply they had heard. Daily released Pipecat as open source in 2024 so developers could reuse that work in programmable media pipelines and swap services as their applications changed. Kwin credits Mark Backman with driving the framework’s development from its launch alongside Daily colleagues and a growing contributor community. Early voice experiments · Pipecat repository · Kwin’s account of Pipecat’s development

What makes a voice agent usable

  • Measure the pause the caller hears. Kwin measures voice-to-voice latency from the end of the caller’s speech to the first audible reply. Microphone processing, network transit, transcription, inference and speech synthesis all count. He uses roughly half a second as a reference for human conversation and recommends aiming for 800 milliseconds in a voice agent. In a 2024 experiment, he ran Deepgram’s transcription and speech generation close to the language model and picked Llama 3 8B after he could not get the 70B version to respond fast enough. He displayed the caller’s wait alongside component timings so developers could see where the time went. 2024 latency experiment · 800-millisecond target
  • Move turn detection into shared infrastructure. Kwin first wrote turn detection in individual applications, then helped generalize it in Pipecat. In his collaboration with Shrestha Basu Mallick, he worked on the application and framework while she worked on models and APIs, where Gemini Live offered its own turn detection. Kwin’s experimental task manager showed what remained for application developers to figure out. He gave the model access to past conversations and sparse instructions for updating the screen. It could display lists on request, yet it cut Shrestha off and put the wrong books on her reading list. The prototype gave him a way to see which instructions the model could infer and where the application needed to spell out the behavior. Gemini Live and Pipecat collaboration
  • Evaluate models across a whole conversation. In 2025, Kwin saw promise in audio-native models because they could preserve speech that transcription might mishandle when a person switches languages. An end-to-end speech model could also reduce the number of inference calls. The models he evaluated at the time were less dependable than leading text pipelines at following instructions and calling tools, so he advised testing on the intended application. His aiewf-eval project exercises long conversations, factual grounding, instruction following and tool use. In his 2026 benchmark, even the models that passed every test were too slow for the workload he was evaluating. 2025 assessment of audio-native models · aiewf-eval · 2026 voice-agent benchmark

Why Daily built Pipecat Cloud

Kwin initially resisted building a hosting product around Pipecat. Daily already ran media infrastructure, and other companies offered platforms for assembling voice agents. But developers kept asking how to deploy and scale their own code without operating Kubernetes. Daily built Pipecat Cloud for that workload. A phone agent has to start quickly enough to answer a ringing call, then stay running for the length of the conversation as demand rises and falls. Nina Kuruvilla and Mark Backman announced general availability in January 2026. Kwin on why Daily built Pipecat Cloud · General availability announcement

Kwin also explored how little infrastructure a voice agent could use. In a joint demonstration with Sean DuBois, a Raspberry Pi inside a stuffed animal connected directly to a laptop over WebRTC on a local network. Kwin wrote the sampler that tried, with mixed results, to make the laptop’s model compose a poem using only two-syllable words. His macOS experiments explore running conversational agents on personal computers, and his writing on WebRTC explains when an agent can use a direct connection without a dedicated media server. Joint WebRTC demonstration · Local macOS voice agents · Direct WebRTC connections

Daily’s acquisition and Pipecat’s governance

Daily and Deepgram were in the same Y Combinator batch, and Deepgram had supplied transcription for Daily’s audio and video APIs since 2019. On October 29, 2026, Kwin announced that Deepgram was acquiring Daily and that he and the Daily team were joining Deepgram. Deepgram said it would sell, support and develop Daily’s products and planned to build its voice-agent platform on Daily’s real-time network and Pipecat. Kwin went from running Daily to leading Deepgram’s open-source work. Kwin’s October 29 announcement · Deepgram’s acquisition announcement · Deepgram leadership

Daily and Deepgram also announced plans to establish an independent Pipecat Foundation with organizations that use and contribute to the project. The proposed steering committee would represent the wider voice AI community, including Deepgram competitors. Under the plan, Pipecat would keep its BSD license and support for competing providers. Kwin said keeping Pipecat open was central to the Daily team’s decision to join Deepgram. The plan lets Deepgram build a commercial platform on Pipecat while developers using the framework can continue to choose their own models, speech services and hosting. Daily’s governance announcement · Deepgram’s foundation plans

What comes after the voice call

In Gradient Bang, a multiplayer game Kwin builds with collaborators, a player can send ships to explore or trade, keep talking while the jobs run, and ask to see their task history on screen. Background agents share context, and the game can compress conversation history without pausing the exchange. Players can also bring their own agents. Kwin compares today’s agents to the web pages he wrote by hand in 1995, which eventually became parts of richer applications. Gradient Bang gives him a place to build what might surround the agent: a world several people can act in, a screen they can change by speaking, and work that continues while their attention moves elsewhere. Gradient Bang · Kwin on bringing agents into the game · The New Primitives

Read the topics behind these talks

5 conference talks

Loading more talks and highlights…

References