Mirror Particle wants brands to stop treating human behavior as a script that can be acted out by a large language model. The two-year-old, San Francisco-based startup is building what co-founder and CEO Abhivyakti Ahuja describes as a world model: an AI system meant to simulate why people act, what changes them, and how those shifts affect what they may do next.
The company is entering a crowded and well-funded category. Simile raised $200 million at a $2 billion valuation; Aaru raised $88 million at a $1 billion valuation; and Humans&, an AI startup that announced a massive $480 million seed round in January at a $4.48 billion valuation, launched Persimmon to model human behavior.
Why Mirror Particle rejects demographic role-play
The common approach to human behavior prediction today leans heavily on LLMs. Those systems are prompted or fine-tuned to act like a target demographic, then used to infer what a group might want, believe, or buy.
Ahuja argues that this is the wrong foundation for the problem. In her view, adding a small layer of fine-tuning on top of models trained on enormous volumes of data does not meaningfully reshape how those models behave. She put it bluntly: "It’s like bringing a super soaker to Niagara Falls."
Her criticism is not only about scale. It is also about what LLMs are built to model. Ahuja says LLMs are centered on written language, while people make decisions through a broader mix of perception, spatial reasoning, social intelligence, and changing context.
That distinction matters for market research and product strategy. If a system mainly reflects what people write or say, it may miss the signals that come from what people actually notice, ignore, choose, avoid, or repeat. For a company trying to anticipate consumer behavior, that gap can change the recommendation entirely.
A world model of the changing person
Mirror Particle says it is building a foundation model from scratch to track behavior as an evolving system. Rather than trying to capture a fixed profile of a consumer, the company is focused on how motivations move over time.
As Ahuja put it, "We don’t want to capture the static person." The goal is to understand longitudinal data: how people are changing, what triggers those changes, and how strongly those triggers matter. If a person or group is not changing, the company treats that as a signal too.
The startup uses a proprietary mix of data to model demographic segments. That mix includes clients' customer data, current events, pop culture, social media, and more. The idea is to see a segment not as a flat persona, but as a population moving through experiences that can alter its motivations.
One core focus is revealed behavior. In plain terms, that means what people do rather than what they report about themselves in surveys. For brands, the difference can be significant: a customer may say one thing in a research setting but behave differently when choosing a product, reacting to packaging, or deciding whether a category matters at all.
What brands could use it for
Mirror Particle is starting in markets where companies already spend money on consumer insight: market research, brand strategy, and product strategy. That gives the startup a practical entry point. It does not need to persuade brands that consumer intelligence matters; it needs to convince them that its model produces better answers.
The company might help a beauty brand evaluate how to reach Gen Z, but Ahuja’s example shows that Mirror Particle wants to go deeper than ad copy. The question may not be how to sell eyeshadow palettes more effectively. It may be whether that demographic wants eyeshadow palettes in the first place, or whether blush is the better product to pursue.
That is the strategic promise: moving from message optimization to decision support. A brand could use the engine to test whether it is asking the right question before it spends money building, positioning, or promoting the wrong thing.
Mirror Particle also emphasizes the "why" behind its predictions. Its engine is designed to give customers the motivations, constraints, and context that support a recommendation. That matters because a prediction without reasoning can be difficult for a brand team to trust, challenge, or act on.
The pet food lesson
One early pilot shows how the model can redirect a brand away from a narrow creative question. A well-known pet food brand wanted to know which image would help packaging sell better. The options included chicken, beef, and vegetables.
Mirror Particle’s technology concluded that the imagery was not the central issue. The brand was already so recognizable that consumers saw it as mass market and cheap. The implication was that sales would plateau until the company addressed that perception problem.
That example captures the broader argument behind the company. A surface-level test might compare packaging visuals. A behavior model, at least as Mirror Particle describes it, should identify when the visible variable is not the real constraint.
The team and the larger ambition
Ahuja’s interest in modeling human behavior is tied to her background in neuroscience and computer science. Originally from India, she studied at the University of Toronto, where she was inspired by Geoffrey Hinton’s contributions to neural networks.
After school, she worked at Amazon Robotics building robots that build other robots. There she met co-founders Will Song and Thomson Yen. Song has worked on sales personalization engines, while Yen has focused on using deep learning to understand how AI agents interpret human behavior.
Mirror Particle has already raised an angel round and says it is close to closing its first venture round. The company is also set to compete in Startup Battlefield 200 at TechCrunch Disrupt 2026 in San Francisco on October 13-15.
The long-term vision is broader than brand research. Mirror Particle wants to become the "general layer for anticipating human behavior" and move from population-level analysis toward individual-level insights. For Ahuja, the larger point is that AI systems and people will need better models of humans if they are going to work alongside each other.