I am a Research Director at Google DeepMind, leading research teams across video understanding, agentic vision, and compact encoders for Gemini. Our native architectures set state-of-the-art benchmarks across core capabilities (see the Gemini 3.0 announcement and Gemini 3 Pro Vision). Most recently, I pioneered agentic video understanding, shifting from static fixed-rate processing to an active loop where the model dynamically searches, scans, and inspects video segments across frames, audio, and transcripts. These models are deployed across YouTube, Search, Vertex AI, and other Google surfaces. Earlier at Google, my research centered on foundational representation learning and novel vision architectures. Prior to joining Google, I completed my PhD at ETH Zurich on efficient data selection, clustering, and discrete optimization algorithms.
Agentic video understanding
Sep 2026A paradigm shift from static video processing (where the model ingests video at a fixed frame rate) to an active, goal-directed agentic loop. Gemini invokes internal tools to dynamically search, scan, and inspect target video segments across visual frames, audio, and transcripts—cutting token usage and overall cost. Read Google post
Token-efficient long-form video analysis (LongVideoBench)
Across long-form video (from 10-minute how-to guides to 90-minute lectures and multi-hour recordings), static processing forces a tradeoff between prohibitive token costs or techniques that drop critical details. Agentic video invokes internal tools to load only relevant video parts, achieving the Pareto frontier of accuracy-to-cost.
Accurate fast action analysis with dynamic FPS
With agentic video understanding, Gemini (e.g. 3.7 Flash) accurately counts fast-paced movements by dynamically scanning and rewatching the video at different frames per second as needed. Unlike static 1 FPS sampling that completely misses rapid motion, the agentic loop isolates sub-second action boundaries.
Token-efficient needle-in-a-haystack search
Using agentic video understanding, Gemini accurately answers complex questions based on video content while consuming a significantly lower number of tokens compared to static analysis. It searches transcripts and audio cues before fetching targeted visual frames.
Gemini live
Sep 2026Gemini 3.8 Live processes visual inputs in near real time (sub-240ms latency across 97 supported languages), enriching conversations with live visual context. Combined with Extended Thinking, the model reasons and speaks simultaneously—handling complex tasks, physical environments, and background tool calls without interrupting live conversation. Read Google post
Agent-to-Agent Live Multimodal Collaboration
Two Gemini agents interacting over live video and audio streams, dynamically coordinating business strategy documents and marketing collateral on screen in real time.
Near real-time visual chess gameplay
Gemini 3.8 Live plays chess in near real-time using visual context from a continuous video feed, evaluating board states and maintaining natural conversational banter without pausing speech.
Live visual onboarding & workspace guidance
Gemini 3.8 Live guides tasks in real time, using continuous visual context from camera feeds to answer live questions as users handle physical equipment, micro-controllers, or workspace tools.
Gemini 3 Pro Vision
Dec 2025A generational jump in multimodal vision and spatial reasoning. Gemini 3 Pro Vision interprets complex documents, follows dynamic events in high frame rate video (up to 10 FPS), localizes visual items with pixel-precise pointing coordinates, and offers granular control over latency and token cost via adaptive media resolution. Read Google post
Pixel-Precise Spatial Pointing & Coordinate Grounding
Gemini 3 Pro Vision moves beyond basic classification to output pixel-precise coordinates for elements in intricate images and diagrams. This unlocks open-vocabulary spatial grounding, UI component localization, and direct visual referencing for agents and downstream tools.
High Frame Rate Video Understanding (up to 10 FPS)
Native support for high frame rate video processing allows Gemini 3 Pro Vision to follow fast actions, identify cause-and-effect sequences across fleeting moments, and perform fine-grained temporal analysis across multi-modal video feeds.
Complex Document & Multi-Page Layout Analysis
Deep reasoning over mixed-format documents, dense technical schematics, tables, and handwritten notes. The model preserves native aspect ratios and scales computation smoothly using adaptive media resolution.