Usama Bin Shafqat

Usama Bin Shafqat is an AI engineer at Google Labs who built the text-to-audio pipeline integrated into NotebookLM’s Audio Overviews. His work turns source material into a two-person conversation rather than a spoken reading of a document. His personal website also credits him with subsequent work on Video Overviews and Slide Decks, and with leading NotebookLM’s evolution into a multimodal knowledge-work agent using a coding harness and virtual machine.[2][1]

Shafqat’s self-described research-workspace contribution extends NotebookLM from explaining sources to performing analysis and creating downloadable artifacts. He coauthored Google’s 2026 research upgrade announcement with Trond Wuellner, director of product management for NotebookLM. The broader system combines Gemini 3.5 and Antigravity with a secure cloud computer that can write and run code, plus more than 100 curated software skills. Users retain control over the sources added to their research.[2][3]

The previous year, Shafqat worked on Video Overviews, a contribution he credits on his portfolio, and coauthored the July 29, 2025 launch announcement with senior UX designer Shan Wang. The team’s initial format used narrated slides, combining newly generated visuals with images, diagrams, quotations and numbers from source documents. Users could specify a topic, learning goal or audience to guide the overview.[2][4]

Those later capabilities followed an audio experiment initially developed outside NotebookLM. In the October 25, 2024 Latent Space interview, Shafqat described a team demo intended to transform sources into engaging, listenable audio. His principal work was the separate text-to-audio engineering pipeline subsequently integrated into NotebookLM. He credited close collaboration with DeepMind’s audio researchers and Gemini 1.5’s ability to absorb long context.[1]

Shafqat explained the two-speaker format as a way for different AI personas to question the material and each other, opening up perspectives beyond a straightforward summary. Natural delivery depended on both layers of the pipeline: the audio model mimicked intonation, breathing, pauses and laughter, while generated text supplied signals about where those behaviors made sense. He also credited Steven Johnson with the suggestion to leave the hosts unnamed.[1]

For Shafqat, an engaging script needed a clear destination, differing perspectives and a gradual release of information. As he put it, “if there's just too much agreement between people, like that's not fun to listen to.” A Wikipedia article might reveal what happened in its opening paragraph; his explanation of the dialogue format instead emphasized setting up tension, exploring a topic and returning to the main narrative. Early experiments included obscure material such as Mountain View city council meeting notes, testing whether transformation could make otherwise overlooked information worth listening to.[1]

The team began with extensive critical listening before formal rater evaluations. Shafqat described improvements to both the generated transcripts and the audio, followed by an evaluation set that broke entertainment into several dimensions rated on a Likert scale. He stressed preserving engagement while addressing invented information, insensitivity and narrative coherence. Listening was therefore part of the engineering process, alongside formal evaluation.[1]

His public engineering advice connected capability exploration with quality work. Adding user controls, he argued, required revisiting the quality process rather than merely exposing another parameter: “The knob actually needs to work reliably.” He encouraged builders to test the edge of a model’s capabilities, seek creative feedback and combine multiple models when one attempt was insufficient. His claim that additional thinking time consistently improved results was an account of his own building experience.[1]

Before his AI product work, Shafqat worked on Google’s data-center supply-chain planning. In the 2024 interview, he recalled wanting more “zero-to-one building” and moving to Area 120, where he worked on a creator-commerce platform. His portfolio names that project Qaya and describes storefronts for selling digital assets with YouTube integration. He traced his subsequent AI work to Area 120’s transition toward Labs. His website also lists a BSE in Computer Science from Princeton University in 2018.[1][2]