Artificial intelligence has moved through several distinct eras in the last five years, and each one has redefined what “building AI” actually means. First, it meant proving a model could solve a commercially meaningful problem. Then it meant building foundation models general enough to survive deployment in the real world. Today it means something less visible but more consequential: making sure a capable model does not fall apart the moment it is asked to act on its own as an agent.
That last problem, gap between benchmark performance and production reliability, has become one of the defining challenges facing frontier AI labs. A model that reasons beautifully in a sandbox can generate a malformed document in the wild. Systems that ace a coding benchmark still falter on the fortieth step of a multi-step task. Closing that gap requires expertise that understands both the commercial and technical layers of AI deeply enough to see where they break.
One technology leader whose career has tracked this exact evolution, one layer at a time, is Juhi Parekh . Her work has spanned applied AI products, foundation models, and post-training infrastructure used to evaluate and improve models.
From Apple to Turing: A Career Built Across the AI Stack
Juhi Parekh is General Manager and Head of Forward Deployed AI Research at Turing, where she works with frontier AI labs on exactly that layer: the post-training pipelines, evaluation benchmarks, and reinforcement learning (RL) environments that determine whether a trained model is reliable enough to ship.
A data scientist turned AI research product leader with an MBA from Northwestern’s Kellogg School of Management, she has spent the last several years working across what she calls three distinct altitudes of AI commercialization, applied AI products at Apple and Samsung Research US, foundation models at Niantic Spatial, and now AI infrastructure at Turing, with earlier machine learning product roles at Amazon and startups.
She speaks regularly on AI product strategy, including talks for Product School’s 150K-member community, podcasts with Fiddler AI, and industry panels alongside leaders from startups like Sierra, and frontier labs.
The Two Constraints That Never Actually Went Away
Parekh’s central observation, developed across her time at Apple, Niantic, and Turing, is that the “hard part” of commercializing AI has not disappeared so much as changed disguises. At every altitude she has worked, two core constraints remain: the technical requirements to build a good model, data, research talent, and computing and the strategic requirement to solve a real-world problem.
“Data, compute and researcher talent density determine what’s possible before strategy decides what’s desirable,” Parekh says. “The bottleneck never truly disappears, it just moves. And understanding why it moves is more useful than any prediction about where AI goes next.
Applied AI: The Bottleneck Was Product-Market Fit
In the applied AI stack, as a product manager working on visual intelligence and search at Apple, and later on large language models at Samsung Research US, Parekh found the model itself was rarely the constraint.
The real question was which of a thousand technically feasible applications were also commercially viable. The work required a venture-capital mindset: identifying user problems that only frontier AI could solve, validating them with market data, and rapidly prototyping to see what survived real-world usage.
Model-building constraints such as data scarcity and infrastructure cost existed, but they were secondary to product-market fit. They surfaced when a model trained on synthetic data looked great in a lab but broke under real conditions, or when infrastructure costs made a promising demo unshippable.
Foundational AI: The Bottleneck Moved to Generalization
By the time Parekh joined Niantic Spatial as research product lead to build the company’s foundation world model, the constraint had shifted. The market had stopped asking “should we build this” and started asking “can the model generalize.” Working on spatial AI, focused on building models that perceive and reason about 3D space — she helped navigate the industry’s transition from narrow, task-specific systems for spatial intelligence toward foundation models capable of generating, perceiving, and navigating the physical world.
The data constraint became an access problem: spatial foundation models need large-scale, real-world 3D data that no one can simply scrape off the internet. Niantic was uniquely positioned here, after a decade of players scanning real-world locations through its games, the company held one of the most valuable geospatial data assets available for training spatial AI.
Meanwhile, the application-identification constraint got harder in the opposite direction. It was no longer about finding a use case for an existing model, but deciding which spatial problems were worth training a foundation model to solve at all. Parekh has published her own analysis of the space and its evolution, including a widely shared market map of spatial AI that charts the field’s move toward generalized intelligence.
“Foundation models are expensive to build and even more expensive to get wrong,” Parekh notes. “Teams that won were not just the ones with clever use cases. They were the ones that recognized which capabilities were about to be commoditized by foundation models.”
Infrastructure: Where Both Constraints Become the Entire Business
Which brings Parekh to where she sits today at Turing, working with frontier AI labs on the layer underneath both of those earlier eras. It is the part of AI that almost never makes it into a keynote, and, in her assessment, the part that currently determines whether anything built on top of it actually works. The bottleneck has moved to post-training.
Here, the data-and-talent constraint is not a supporting cost center, but the product itself. Instead of lacking compute or model architecture ideas, frontier labs now lack the specific, high-signal private data needed to make post-training work, long-horizon expert-authored reasoning tasks, PhD-level rubrics that let a model’s output be graded correctly.
“Talent in this context means domain experts who can write a task hard enough that a frontier model actually fails at it,” Parekh says. “That’s a much scarcer skill than it sounds.”
The application-identification constraint shows up in a form Parekh did not initially expect: the focus is no longer just finding a market but identifying specific failure modes worth building a benchmark or RL environment around. A model can pass every standard evaluation and still fail as an agent by mis-formatting tool calls or confidently reporting a task as finished when it is not, simply because nobody identified that scenario as worth testing.
Some of that work is publicly visible. Parekh’s team at Turing has released expert-authored evaluation datasets on Hugging Face, including Multimodal-STEM-HLE++ — a benchmark of PhD-level STEM problems that require models to reason over complex scientific text and images spanning electromagnetism, fluid dynamics, topology, and graph theory,and Rubric-Graded Advanced PhD Reasoning dataset, a dataset of expert-written rubrics designed so that a model’s multi-step reasoning can be graded rigorously. Both reflect the design philosophy behind her work: if a frontier model can pass a test, the test is not doing its job.
“Designing a test problem hard enough that a capable model actually fails at it, then feeding that failure back into training, is now one of the highest-leverage things a frontier lab can do,” Parekh argues.
The Throughline
None of these eras replaced what came before it, Parekh notes. Product-market fit still matters, and foundation models still need someone to decide what is worth building. What changed is where the constraint sits, and every time it moves, it moves up a layer of abstraction: from the application to the model, to the infrastructure that trains and evaluates the model.
As the AI industry moves from releasing capable models to deploying autonomous agents, professionals who understand the full stack, from commercial validation to model generalization to post-training infrastructure, will play an increasingly central role.
“Don’t assume the bottleneck you learned to solve is still the one worth solving,” Parekh reflects. “The AI industry has relocated the hard part at least twice in five years. Betting on where it lands next is a lot more useful than getting good at where it used to be.
Juhi Parekh is General Manager and Head of Forward Deployed AI Research at Turing, where she works with frontier AI labs on post-training pipelines, benchmarks, and RL environments. She previously held product roles at Apple, Amazon, Niantic Spatial, and Samsung Research US.
