Nidhi Kaushik Vyas

Nidhi Kaushik Vyas is an AI product professional whose work connects language and multimodal research with tools that help people find information, express preferences, and make decisions. She led Gemini Embedding 2 at Vertex AI before returning to Google DeepMind to work on visual intelligence and agentic commerce, extending a career that spans low-resource language systems, training-data quality, and AI-generated image watermarking.

From language technology to multimodal products

Vyas earned a master’s degree in Language Technologies at Carnegie Mellon University in 2019. She co-authored the ARIEL-CMU low-resource language systems report, a collaborative effort covering machine translation, entity discovery and linking, and the detection of situation information in text and speech. These systems had to extract useful information where language resources were limited.

Her subsequent research addressed the quality of the examples models learn from. With Shreyas Saxena and Dennis DeCoste, she co-authored work on data-dependent dynamic learning rates, which adjusts learning rates for individual training examples so an optimizer can reduce the influence of corrupted data. Her Galileo-affiliated research also explored repairing datasets to improve models, combining the identification of faulty examples with the selection of informative samples. She was among the original contributors to Galileo’s Data Error Potential metric, designed to help practitioners locate difficult training data and prioritize investigation.

Vyas then helped turn AI research into usable products. She and Zahra Ahmed drove product delivery for the initial SynthID image-watermarking release in 2023, working alongside its research and engineering team. SynthID embeds an imperceptible watermark into generated image pixels, enabling detection when metadata is lost and after common image transformations. Her contribution was product delivery within a collaborative effort, rather than sole invention of the watermarking technology.

At Vertex AI, Vyas led Gemini Embedding 2 before returning to DeepMind to focus on visual intelligence and agentic commerce. She also co-authored the Gemini Embedding 2 technical report. The model places text, images, video, audio, and documents into a shared representation space, allowing retrieval to connect information across different kinds of content. That work provides a useful bridge to her commerce interests: a person’s reference image and written request can both carry essential information about what they want.

Helping people develop an intention

Vyas’s approach to shopping agents starts with an articulation gap. People often recognize what appeals to them before they know how to describe it. A shopper may have a photograph, a budget, and a feeling about how a room should look, yet lack the vocabulary or product knowledge needed to turn that preference into a search query. She argues that an agent should help develop the intention as well as execute it.

Her multimodal commerce framework organizes that collaboration into discovery, research, and response. A living-room redesign supplies the concrete example: the agent must learn what furniture will fit, which aspects of a reference image matter, and how those preferences connect to products that are actually available. Each stage has a different job, and each needs its own evaluation.

• Ask the question that changes the decision. During discovery, the agent builds a working state from conversation history, explicit requirements, reference images, and relevant context. It separates hard constraints, such as budget, from inferred preferences, such as style, and tracks confidence in those inferences. It also identifies changing information, including inventory, that needs refreshing. Vyas uses expected information gain to prioritize clarification: room width may matter more than a precise style label because furniture that does not fit cannot satisfy the request. Repetitive or unnecessary questions count as failures, even when they gather more information.

• Use visual preference elicitation. When a shopper cannot describe a style, the agent can offer a visual board with options related to their reference image. The examples give the person something concrete to respond to, while selections, clicks, and other interaction signals help update the agent’s confidence. Making those preferences useful for retrieval requires connecting them to product-catalog attributes and the catalog’s ontology—its organization of product concepts and relationships. Vyas gives merchants’ domain expertise an important role in building that connection.

• Treat response format as part of intelligence. A product-policy question may need a short summary; a comparison may need a table organized around the attributes the shopper cares about; a room redesign may need images and inspiration. Vyas treats the choice of presentation as part of the agent’s task because useful information can still fail when it is difficult to find or act on. In the system she describes, merchants contribute product knowledge while the agent controls presentation, maintaining a consistent experience across merchants.

• Evaluate the whole collaboration. Vyas tests whether the working state retains facts, calibrates confidence, and responds appropriately to changed inputs. Altering one requirement should update the corresponding constraint without disturbing unrelated ones. The questioning strategy must identify blockers, ask useful questions, and avoid over-asking. User simulators with hidden preferences help assess how efficiently an agent uncovers those preferences and whether it chooses the right elicitation format. Final responses are checked for factual fidelity, presentation, and actionability. These evaluators should develop with the product as its behavior becomes more complex.

Vyas also preserves a meaningful role for the person doing the shopping. In the user studies she describes, people wanted to participate directly in discovery and inspiration: exploring possibilities was part of deciding what they liked. Later tasks, including price comparison and negotiation across merchants, appeared more suitable for delegation. At the time of the presentation, her team had not yet reached agent-to-agent interaction. Her framework assigns the agent substantial research and comparison work while keeping the shopper involved in forming the preference that guides it.