What It Takes to Make Complex AI Systems Usable Across Teams

How Doordash’s Nam Pham and Their Team Turned AI Research Skills Into Shared Infrastructure for Research, Design, and Product

Kate Towsey

Jul 22, 2026

Nam Pham is a senior UX researcher at DoorDash, where they use mixed-methods research to build zero-to-one products. Lately, Nam’s been building dining out as a new category and shaping affordable DoorDash dining. They’re also developing research evaluation practices, making sure that users—in all their messy, human complexity—stay front and centre as the team builds AI and large language model (LLM) products. Before DoorDash, Nam led research on homeowner and seller products at Realtor.com. They studied at Parsons School of Design, where they developed a deep belief in participatory design.

How to AI UXR is supported by Strella, an AI-powered customer research platform that partners with you to build, moderate, and synthesise interviews, allowing you to go from question to actionable insights in just a few hours.

In This Conversation

AI research systems are becoming increasingly easy to build, but how do you ensure those systems are reliable, properly maintained, and deployable beyond the person who built them and their local drive? In this episode, Nam Pham shares a powerful AI research system—one of the most advanced I’ve seen—that’s now part of DoorDash’s core UX infrastructure.

In this episode, Nam demos their research workflow in Cursor and a study plan for a specific delivery question. This system helps DoorDash researchers, designers, and product managers move through the research process, from scoping and planning through instrument design, survey programming, analysis, and sharing.

We talk about how Nam and their DoorDash colleagues use a more traditional tool stack, combined with skills, harnesses, hooks, scripts, connectors, and an internal “skills marketplace” to share and update AI skills that help teams do research.

The primary takeaway from this episode is this: AI might enable you to build solo, but building in collaboration with engineering, design, analytics, and others will enable you to deliver systems that operate beyond one person and their local machine.

The How to AI UXR Map

This series builds on the insights shared in the How to AI UXR map, a five-page map that charts key trends, helps you pinpoint your AI maturity level, and offers practical, real-world applications you can adapt to your research systems.

Download the Map

In this episode, we cover:

How DoorDash turned individual AI experiments into shared research infrastructure
What it takes to make an agentic research system usable by researchers, designers, and product managers
Why a useful AI skill is closer to an operating procedure than a prompt
How internal knowledge from research repositories, company systems, and Slack can shape research planning
Why Markdown, reference files, Python scripts, and deterministic code help make LLM outputs more reliable
How AI can programme a Qualtrics survey while the researcher keeps working, and where human review still matters
How an internal “skills marketplace” for sharing and updating AI skills supports adoption, maintenance, and shared standards
How collaboration with engineering, design, analytics, and quantitative research changed what was possible
Why Nam argues for patience: teach the agent, test it with colleagues, observe where it fails, and continuously and collaboratively improve the system

Partway through the episode, as a leader building one of the tools these researchers are exploring and using, Priya Krishnan, the cofounder and COO of Strella, shares her take on the conversation.

Things Referenced

Cursor, an AI coding agent
Claude Code, another AI coding agent
Glean, an AI platform for work
MCP (Model Context Protocol) connectors are universal, plug-and-play software bridges that allow AI agents to safely access external tools, databases, and third-party applications
Markdown is a lightweight plain-text formatting language that uses simple symbols like #, *, and - to structure documents. It’s relevant to AI because it bridges human intent and machine processing, allowing LLMs to easily understand, organise, and output complex information without wasting computing power.
Python is a computer programming language that’s used to give clear instructions to computers. It acts like a translator, turning English-like words into a format that a machine can easily understand.
SQL (Structured Query Language) is the standardised programming language used to communicate with, manage, and retrieve data from relational databases.
AI harnesses are the surrounding software infrastructure, such as tool execution, memory, and safety guardrails, that safely control an AI model and enable it to autonomously execute multi-step tasks.
AI coding hooks are automated, user-defined scripts that trigger at specific points in an AI assistant’s workflow to enforce guardrails, run tests, or format code.
Agent journaling is the process by which an AI coding agent maintains a continuous, detailed log of its internal reasoning, tool executions, and step-by-step progress to help developers audit, debug, and understand its decision-making and workflow.
Skill chaining is the process of an AI agent sequentially linking multiple specific capabilities or tools together, using the output of one action as the input for the next to accomplish a complex, multi-step goal.