Shaped: How Real-Time AI Retrieval Powers Search & Feeds

In an era where personalization defines user experience, Shaped positions itself at the very core of how digital products retrieve, rank, and serve information in real time. Founded in 2021 and headquartered in New York, Shaped is building a real-time retrieval engine designed for search, recommendation feeds, and AI agents. The company operates at the intersection of machine learning, data infrastructure, and user experience, addressing a problem that nearly every modern product team faces: how to make relevance fast, scalable, and maintainable.

As AI-driven products evolve—from personalized feeds to autonomous agents—the need for intelligent retrieval has grown exponentially. Shaped focuses on the infrastructure layer that sits between raw data and user-facing intelligence. Instead of treating retrieval, ranking, and personalization as fragmented systems, Shaped unifies them into a single, coherent interface that works at a production scale.

Backed by a growing team of 25 and supported by primary partner Brad Flora, Shaped remains an active company following its participation in the Winter 2022 batch. Its mission is clear: simplify how teams build relevance into their products without sacrificing performance, flexibility, or speed.

Who Is Behind Shaped and What Experience Drives the Company Forward?

At the center of Shaped is its founder and CEO, Tullie Murrell, a technologist with deep roots in artificial intelligence research. Before founding Shaped, Murrell worked as an AI researcher at Facebook AI Research (FAIR), one of the most influential research labs in the field. His background reflects not only academic rigor but also hands-on experience with systems that operate at a massive scale.

Murrell is also widely recognized as the creator of PyTorchVideo and an original core contributor to PyTorch Lightning—two tools that have become foundational in the machine learning ecosystem. This history underscores a recurring theme in Shaped’s DNA: building practical, developer-friendly abstractions that translate complex machine learning workflows into accessible tools.

Driven by a passion for AI, machine learning, and statistics, Murrell founded Shaped with a clear understanding of where existing tooling falls short. His experience observing how real-world ML systems are deployed—and where they break—directly informed Shaped’s product philosophy: powerful systems should not require brittle architectures or excessive glue code to function.

What Core Problem Is Shaped Trying to Solve for Engineering Teams?

Despite major advances in vector databases and embedding-based search, relevance remains one of the hardest unsolved problems in production AI systems. According to Shaped, most engineering teams today are trapped in what can best be described as a “Frankenstein” architecture—a tangled stack of tools that were never designed to work seamlessly together.

To build something as seemingly simple as a personalized “For You” feed or to give an AI agent contextual memory, teams often need to combine multiple components: a vector database for retrieval, a feature store such as Redis, a separate reranking service, and thousands of lines of custom Python code to glue everything together. Over time, this stack becomes fragile, difficult to debug, and expensive to maintain.

The result is not just technical debt, but also slower product iteration. Simple feature requests—such as adding cart upsells or conversational recommendations—can take months instead of days. Shaped identified this pain point as systemic, not incidental, and set out to re-architect how relevance pipelines are defined and executed.

How Does ShapedQL Redefine Retrieval, Ranking, and Relevance?

Shaped’s answer to this complexity is ShapedQL, a domain-specific SQL dialect purpose-built for search, feeds, and AI agents. Rather than forcing engineers to orchestrate multiple services manually, ShapedQL provides a single declarative interface that compiles down to a high-performance, multi-stage ranking pipeline.

The insight behind ShapedQL is simple but powerful: while retrieval has become easier thanks to vector databases, ranking and relevance logic remain deeply fragmented. ShapedQL brings these pieces together by allowing teams to define the full relevance lifecycle in one query, using concepts engineers already understand.

This SQL-like approach does not sacrifice performance for simplicity. Instead, ShapedQL translates high-level intent into optimized, real-time pipelines capable of serving results in milliseconds. It allows teams to think in terms of outcomes—what should be retrieved, filtered, scored, and reordered—rather than implementation details.

What Are the Four Stages of Modern Relevance According to Shaped?

At the heart of ShapedQL is a structured view of relevance as a four-stage process. Each stage reflects a critical step in delivering meaningful results to users or AI agents.

The first stage is retrieval. ShapedQL can fetch candidates from multiple sources simultaneously, including hybrid search, collaborative filtering, and trending signals. This ensures that the candidate pool is broad, diverse, and contextually aware.

The second stage is filtering. Here, hard constraints are applied to remove irrelevant or invalid results. These might include conditions such as product availability, geographic distance, or business-specific rules that must be enforced before ranking.

The third stage is scoring. This is where machine learning models come into play. Results are ranked using real-time ML models optimized for specific objectives, such as clicks, purchases, or watch time. Because scoring is integrated directly into the query pipeline, it can respond dynamically to user or session context.

The final stage is reordering. Rather than showing users the same items repeatedly, ShapedQL enforces diversity constraints that improve exploration and long-term engagement. This step is especially critical for feeds and agents, where repetition can quickly degrade user experience.

How Does ShapedQL Reduce Engineering Complexity and Maintenance Costs?

One of the most compelling outcomes reported by Shaped is the dramatic reduction in maintenance code. Teams that previously managed over 2,000 lines of custom logic have been able to replace that complexity with approximately 30 lines of ShapedQL.

This reduction is not merely cosmetic. By centralizing logic into a single declarative query, teams reduce the surface area for bugs, eliminate synchronization issues between services, and make relevance logic easier to reason about. New engineers can understand and modify pipelines without reverse-engineering a web of interconnected scripts.

The impact on delivery speed is equally significant. Features that once required months of planning, coordination, and implementation can now be shipped in days. For product teams operating in competitive markets, this speed can translate directly into increased engagement and revenue.

How Does Shaped Support Real-Time AI Agents and Contextual Intelligence?

As AI agents become more prevalent, the need for contextual memory and real-time decision-making has intensified. Shaped is explicitly designed to support these use cases by allowing queries to incorporate text, user context, and session data in real time.

This capability enables AI agents to retrieve information that is not only relevant in isolation but also personalized to the current interaction. Whether an agent is recommending products, answering questions, or navigating workflows, ShapedQL ensures that retrieval and ranking adapt instantly to changing context.

By treating agents, feeds, and search as variations of the same underlying problem, Shaped avoids siloed solutions and instead provides a unified foundation for intelligent systems.

What Integration Options Does Shaped Offer Beyond SQL?

While ShapedQL is central to the platform, Shaped recognizes that not every team prefers to work directly with SQL. To accommodate different development styles and production requirements, the company also offers Python and TypeScript SDKs.

These SDKs allow teams to integrate Shaped into real-time production environments while preserving the same underlying capabilities. Developers can choose the interface that best fits their stack without compromising performance or functionality.

This flexibility reinforces Shaped’s broader philosophy: powerful infrastructure should adapt to teams, not force teams to adapt to it.

How Has Shaped Demonstrated Real-World Impact Through Case Studies?

Shaped points to real-world case studies that highlight its impact on leading brands. By improving relevance, reducing latency, and accelerating feature development, Shaped has helped organizations drive measurable gains in both engagement and revenue.

These case studies illustrate that Shaped is not an experimental research project, but a production-ready system already delivering value at scale. The ability to explore these examples publicly reinforces confidence in the platform’s maturity and applicability across industries.

Why Does Shaped Represent a Shift in How Relevance Infrastructure Is Built?

Shaped represents a broader shift in thinking about AI infrastructure. Instead of treating retrieval, ranking, and personalization as separate problems, it frames them as components of a single, cohesive system. By abstracting complexity without hiding power, Shaped enables teams to build sophisticated relevance pipelines with clarity and speed.

As digital products increasingly rely on real-time intelligence, platforms like Shaped may define the next generation of search, feeds, and AI agents. Its combination of deep ML expertise, developer-friendly design, and production focus positions it as a foundational layer for the future of personalized experiences.

In a landscape crowded with tools that solve isolated problems, Shaped stands out by solving the whole problem—end to end, in milliseconds.