Kwindla Hultman Kramer

Kwindla Hultman Kramer is a cofounder and CEO of Daily and works on the Pipecat open-source framework. His AI work spans conversational infrastructure, model orchestration, and evaluation of voice agents. In his public engineering writing, he treats response speed, reliable tool use, and conversation context as connected requirements: a model that answers correctly can still be unsuitable for a live conversation if it takes too long to respond.[3][6]

On February 2, 2026, Kramer announced Daily’s open-source voice-agent benchmark, linking the implementation at kwindla/aiewf-eval. It tests tool calling, instruction following, and factual grounding across 30 scripted conversation turns, using the same prompt for each model. The scenarios include updating customer records and creating support tickets—tasks where an agent’s words must match its actions. The repository credits Modal for compute resources and Charles Frye for advice about models and inference tuning.[6][7]

Kramer reported that, in the February tests, no model both achieved perfect benchmark performance and responded fast enough for the team’s voice-agent requirements. He used a voice-to-voice target below 1,500 milliseconds, translating to roughly 700 milliseconds to the first token for a text-model pipeline. These measure different stages: text-model latency starts with the inference request, while speech-to-speech latency runs from the end of the user’s speech to the beginning of the model’s audible reply. His production approach therefore combined model selection with model-specific prompting and context engineering.[6]

One consequential scoring choice was to penalize agents that ask for information the user has already supplied. Kramer’s example is an agent that knows a user’s name but asks for it again before submitting a vote. He considers this an instruction-following failure because it interrupts the workflow and annoys the user. By contrast, the benchmark does not penalize whether a model speaks alongside a tool call: models have different defaults, and developers disagree about the desired behavior. The judging system uses Claude Opus 4.5 through the Claude Agent SDK, with the team scrutinizing both raw responses and judging output.[6]

Kramer attaches limits to those comparisons. The scripted inputs continue regardless of a model’s response, whereas real users can adapt; a shared prompt also leaves out the model-specific tuning used in production. API variability further affects repeatability. As he puts it in the benchmark announcement, “A benchmark is a single data point, not a comprehensive judgment.”[6]

In August 2024, Kramer announced the team’s Daily Bots hosted platform and explained the division between Pipecat’s server-side AI orchestration and RTVI’s real-time inference clients. Orchestration covered prompting, state management, data flow between models, and calls to external systems. He described both projects as vendor neutral and made portability an explicit part of the hosting proposition: “If your needs evolve beyond Daily Bots, you can take your code to another platform or stand up your own infrastructure.” The announcement also described word-level interruption handling and phrase endpointing that combined voice activity detection, semantic cues, and noise-level averaging. The page subsequently carried a notice that Daily Bots was merging with Pipecat Cloud.[5]

Kramer advocates combining models and inference loops within an agent. In a retrospective account, he described a January 2024 team-built Pipecat agent that paired the small, fast Moondream vision model with GPT-4 so a voice agent could process its surroundings at a high frame rate. His February 2026 examples included specialized models processing video, checking content safety, and compacting context alongside the main conversation. He argues that context engineering will remain necessary as applications become more ambitious: “My rule of thumb is: we will always want to do 20% more than the best models can (easily) do.”[6]

In his July 2025 WebRTC guidance, Kramer recommended direct, “serverless” connections for most teams managing their own voice-agent infrastructure. Here, serverless means connecting the client directly to the cloud-hosted agent, eliminating a separate media-server hop and its operational burden. His recommendation is conditional: a geographically distributed commercial WebRTC cloud can improve average latency through private mesh routing and provide extensively tested reliability features. Direct connections cover one human and one agent; multiparty sessions require servers, and he generally recommends servers for video.[8]

Kramer has also taught voice-agent engineering publicly. Latent Space’s June 2025 AI Engineer World’s Fair preview listed his Gemini and Pipecat workshop, joint sessions with Shrestha Basu Mallick, and a realtime-AI keynote. The publisher also released a Voice AI Masterclass with Kramer and swyx.[1][2]

Before his voice-AI work, Kramer worked on networked systems and collaborative computing. He traces his interest in large-scale networks and real-time video to graduate study with Mitchel Resnick at the MIT Media Lab. Before Daily, he helped found Oblong Industries, which built an operating system for spatial, multi-user, multi-screen, multi-device computing and a collaborative big-data environment.[3]

Earlier still, Kramer helped build AllAfrica’s technology, spending six years working mostly from its Washington, DC office. In his personal account, he recalls advocating free and open-source software for large-scale web development, particularly in public and civil-sector organizations during the early 2000s. His argument favored spending on education and training rather than Windows licenses. His experimental networked-device work also included computational tiles with infrared communication and a bytecode interpreter; he credits Nelson Minar as his collaborator on mobile-code simulations.[3]