Raiza Martin

Raiza Martin led the NotebookLM team at Google Labs, helping shape an AI research assistant grounded in users’ own documents and its conversational Audio Overviews. Her work centered on observing how people study and research, choosing a focused product experience, and using direct feedback to refine it. She later left Google and cofounded Huxe, a personal audio companion whose team announced its wind-down in May 2026.[1][4][5][6]

Martin and editorial director Steven Johnson jointly announced NotebookLM on July 12, 2023, when Project Tailwind began rolling out under its new name. The initial product used selected Google Docs to generate summaries, answer questions and suggest new connections. Source grounding distinguished it from a general chatbot: answers drew on the user’s material, with citations and original quotations to help check them. The announcement said grounding appeared to reduce hallucination risk while still requiring users to fact-check responses against their sources.[3]

Martin studied Johnson’s research practice through both interviews and observation. Watching him work revealed details that interviews alone did not capture: how he organized notes, questioned his own reasoning and pushed the developing product to its limits. In the October 2024 Latent Space interview, she recalled thinking, “maybe Steven is the product.” Her proposed task was to make the benefits of his expert workflow accessible to everyday users, then test whether those users actually wanted them.[1]

For Audio Overviews, Martin argued that reading a document aloud was insufficient. Engaging listening required an editorial transformation: two personas could take different angles on the material and bring it to life through dialogue. AI engineer Usama Bin Shafqat described building the text-to-audio pipeline separately before integrating it into NotebookLM, combining Gemini 1.5’s long-context capabilities with DeepMind’s audio work. He explained that natural breathing, pauses and laughter involved both the audio model and signals supplied during text generation.[1]

Martin said the team prioritized audio over another creation-interface direction, asking where it could deliver distinctive value. She also resisted adding controls simply because users requested them. Mocking up more settings made her reconsider what people liked about the original experience: “That they just pushed a button and it was cool.” Additional controls, she argued, needed to preserve that ease. A wider range of outputs also made quality harder to measure.[1]

Her approach to evaluation began with an opinionated quality bar. Martin advocated listening, rejecting weak results and iterating until the team found the output impressive, then using evaluations to validate that judgment. Bin Shafqat described team listening sessions followed by formal rater evaluations across several dimensions, using a Likert scale. Entertainment and engagement remained central alongside groundedness, narrative coherence, hallucination checks and safety.[1]

Martin wanted to hear from users immediately after they tried a feature and ask follow-up questions. NotebookLM’s Discord community gave her a way to investigate use cases, failures and the reasons people persisted despite difficulties. She reported that community messages sometimes exposed service problems faster than internal monitoring. Her questions extended beyond initial enthusiasm to whether a tool remained useful tomorrow or the following week, and why people kept using it or stopped.[1]

That feedback also informed decisions to simplify. In the October 2024 interview, Martin described a feature that let users highlight source text and transform it. She said it was little used, architecturally complicated and difficult to maintain as new features arrived. The team had begun a “50-50 sunset” experiment to see whether users complained; she reported none so far. The experiment reflected her preference for a strong core feature set and the ability to withdraw features.[1]

Martin framed AI product design around the desired outcome: delighting someone, simplifying a workflow or delivering a high-quality result. She called the work “a craft as much as it is a science” and argued that builders needed a strong view of the value they wanted to deliver rather than getting lost in model capabilities. Her advice to product managers was practical: “Build things yourself.” Testing tools and APIs, she said, helped establish the conviction behind a proposed product.[1]

NotebookLM’s development began before its 2023 announcement. Looking back in the October 2024 interview, Martin described earlier work in startups, advertising and payments, followed by leadership of AI Test Kitchen. In October 2022, while considering products beyond technology demonstrations, she encountered an engineer named Adam’s project, Talk to Small Corpus, which used a language model to let people converse with their data. She characterized the early approach as essentially retrieval-augmented generation, or RAG.[1]

Martin connected that prototype to her experience attending college while working full time: being able to talk to a textbook could have helped when she was tired after work. She showed prototypes to college students, particularly adult learners, and recalled that they understood the usefulness without much explanation. The team iterated toward the 2023 Google I/O demonstration. In retrospect, she wished they had launched sooner so they could learn faster.[1]

On December 2, 2024, Martin announced her departure from Google after 5.5 years, most recently leading NotebookLM. She was beginning a venture with Jason Spielman and Stephen Hughes. She described more than four years of collaboration with Spielman across teams, including hours of shared-screen product design in Figma, and praised Hughes’s product-minded, user-centric engineering. The three, she wrote, had each wanted to join a startup and saw an opportunity to build something transformative.[4]

Martin subsequently cofounded Huxe, described by O’Reilly as a personal audio companion that turned subjects users cared about into generated interactive experiences. The AI Engineer World’s Fair June 2025 recap listed a Huxe launch alongside Martin’s AI Product Management keynote. Huxe’s website described audio that users could interrupt to request a different explanation, more technical detail or a change of subject.[5][2][6]

On May 21, 2026, the Huxe team announced that it would wind down the product and stop development. The announcement scheduled app-store removal that day, service ending on May 28 and permanent deletion of user data on May 29. It said the team was moving on to new things.[6]