Mariya Mansurova on Product Analytics & Agentic AI's Impact
Mariya Mansurova, a Product Analytics Manager, brings over 12 years of experience in product analytics across diverse industries, from search engines to fintech. Her career path, rooted in software engineering, mathematics, and physics, and complemented by hands-on experience as a product manager, offers a comprehensive perspective on how analytical teams can empower businesses to make informed decisions. Mansurova is driven by the pursuit of novel insights and innovative methodologies, reflecting a deep-seated belief that curiosity fuels progress.
A significant area of her focus is the emerging field of agentic AI. Initially drawn by curiosity and the pervasive discussion around large language models (LLMs), Mansurova quickly recognized the transformative potential of agentic systems. She believes their influence on daily life and professional workflows will only intensify. Her practical engagement spans utilizing tools like GitHub Copilot and Claude Desktop to building custom agents with frameworks such as smolagents, LangGraph, and CrewAI. For Mansurova, coding stands out as the most impactful application of agentic AI. While recent research has debated the exact efficiency gains, she personally observes a notable difference, estimating a 20% increase in speed, particularly for repetitive tasks (e.g., SQL pivoting) or when grappling with unfamiliar technologies (e.g., building a web app in TypeScript). This boost, she suggests, represents a paradigm shift, expanding the realm of what feels achievable and potentially creating an efficiency gap between those who leverage these technologies and those who do not. Looking ahead, she is particularly enthusiastic about automatic reporting agents, envisioning AI capable of data retrieval, visualization, root cause analysis, and even drafting presentations—a vision she has prototyped for KPI narratives.
Mansurova is also a strong advocate for computer simulations in product analytics, a tool she believes remains significantly underutilized. Her “Practical Computer Simulations for Product Analysts” series aims to demonstrate the power and accessibility of this approach. Simulations provide a quantitative and accurate method to answer “what-if” questions—such as estimating operational agent needs for a new control or predicting the impact of a feature launch in a new market—even when hard data is unavailable. She highlights the efficacy of simulations in navigating uncertainty and distributions, often preferring bootstrap methods to complex statistical formulas. The advent of modern computing power, enabling thousands of simulations in mere minutes, has revolutionized problem-solving capabilities for analysts.
When transitioning LLM applications from prototype to production, Mansurova observes a common pitfall: underestimating the substantial gap between the two phases. Prototypes, while excellent for proving feasibility and generating excitement, offer no inherent guarantees regarding consistency, quality, or safety in diverse, real-world scenarios. Successful production deployment, she emphasizes, hinges on rigorous evaluation. This includes defining clear performance metrics (e.g., accuracy, tone, speed) and continuously tracking them throughout iteration. Analogizing to software development, she asserts that LLM applications demand the same systematic testing. This is particularly crucial in regulated environments like fintech or healthcare, where reliability must be demonstrated to both internal teams and compliance stakeholders, often requiring extensive development time for monitoring, human-in-the-loop processes, and audit trails.
Mansurova’s work frequently integrates engineering principles with data science and analytics best practices, reflecting a belief that the lines between data and engineering are increasingly blurring. She contends that today’s data analysts and scientists require a multidisciplinary skill set encompassing coding, product management, statistics, communication, and visualization. Her early background in programming significantly enhanced her analytical efficiency, fostered better collaboration with engineers, and enabled her to build scalable, reliable solutions. She strongly encourages analysts to adopt software engineering best practices like version control, testing, and code review to improve process reliability and deliver higher-quality results.
Adopting a broad, problem-centric approach, Mansurova views all analytical tools—from statistical methods to modern machine learning techniques—as part of a single toolkit. Echoing Robert Heinlein’s sentiment that “specialization is for insects,” she describes analysts as “data wizards” who select the most fitting tools to solve problems, whether it’s building an LLM-powered classifier or utilizing causal inference for strategic decisions. This mindset, she argues, not only leads to superior outcomes but also cultivates a continuous learning culture essential in the rapidly evolving data industry.
Her prolific writing, which spans topics from text embeddings to multi-AI agents, is characterized by its cohesiveness and accessibility. Mansurova typically writes about subjects that currently excite her, drawing inspiration from new learnings, discussions, online courses, books, and daily tasks. She consistently considers her audience, aiming to create content that is genuinely helpful, both for others and for her future self. Her blog serves as a personal knowledge base, with articles often referencing each other to illustrate the interconnectedness of data concepts.
Mansurova’s structured approach to writing complex topics is deeply ingrained. She often adopts a “concept-first” communication style, starting with foundational principles and iteratively moving towards conclusions. When engaging with online courses, she concurrently outlines the structure, noting nuances and areas for further exploration. A key part of her process involves applying new knowledge to practical examples, as she believes true understanding emerges only when encountering real-world edge cases and friction points. Her writing process involves two distinct phases: an initial drafting stage focused on capturing ideas and code, followed by a meticulous editing phase to refine structure, visuals, and key takeaways. Final reviews, including feedback from her partner, ensure comprehensiveness and accessibility.