Dropbox Innovation & Technology Culture

What project are you most excited to work on in 2025, and what is particularly compelling about this work for you?

In 2025, I am most excited to work on the new AI feature for Dropbox Dash. This project is particularly compelling because it presents an opportunity to integrate some of the most cutting-edge AI capabilities into real-world applications, directly impacting how users interact with their data. The ability to merge state-of-the-art AI advancements with Dropbox Dash is both exciting and challenging.

What does the roadmap for this project look like? Who will you collaborate with, and what challenges will you need to overcome in the process?

The roadmap for this project involves several key phases. Initially, the focus will be on defining the AI feature’s scope and identifying the most relevant use cases for Dropbox Dash users. Next, we will focus on prototyping, iterating and testing the AI capabilities to ensure they integrate smoothly with the platform. As we move into the implementation phase, I’ll collaborate closely with engineers, product managers and UX/UI designers from different teams to ensure the feature aligns with both technical requirements and user needs. 

One of the key challenges I anticipate is ensuring that the AI remains simple and intuitive while leveraging its full potential. I plan to overcome this by maintaining a strong focus on user testing and feedback to iterate quickly and fine-tune the feature. Additionally, addressing potential scalability issues early will be critical to smooth deployment.

What in your past projects, education or work history best prepares you to tackle this project? What do you hope to learn from this work to apply in the future?

My background in machine learning research during my PhD, combined with my experience as a machine learning engineer at Dropbox during my three internships, has uniquely prepared me for this project. Throughout my internships, I worked on real-world ML applications, honing my ability to build, optimize and scale machine learning models effectively. These experiences, along with my deep understanding of AI and its practical applications, have equipped me with the technical skills needed to integrate AI into Dropbox Dash. I also gained valuable experience in cross-functional collaboration, which will be crucial for working efficiently with other teams. Through this project, I hope to deepen my expertise in deploying AI at scale and learn how to evaluate and fine-tune AI solutions for enterprise use cases.

What’s your rule for releasing fast without chaos — and what KPI proves it?

Building Dropbox Dash taught us that releasing fast only works when evaluation is built in. We treat every change, from prompts to retrievers to model settings, with the same rigor as production code. Each pull request runs about 150 canonical queries, judged automatically in under ten minutes. Metrics like source F1 (≥ 0.85) and latency (p95 ≤ 5 s) keep us accountable. This structure enables fast, confident releases across a platform trusted by more than 700 million users.

What standard or metric defines “quality” in your toolchain?

At Dropbox, quality is measurable, versioned, and enforced. Every change is scored across Boolean gates like “Citations present?”, scalar budgets such as source F1 and latency, and rubric scores for tone, clarity, and formatting. We use LLMs as judges, guided by calibrated rubrics that check factual accuracy and context alignment. The results feed shared dashboards so quality stays visible, repeatable, and reliable across Dropbox’s global infrastructure.

Share one recent adoption and its measurable impact.

One of our most impactful adoptions in building Dropbox Dash has been using LLMs to evaluate LLMs. Instead of static BLEU or ROUGE scores, we build judge models that grade factual accuracy, citation correctness, and clarity. This automation keeps evaluation continuous and scalable. Each change is tested and verified before release, backed by rigorous datasets, actionable metrics, and automated gates. It’s how Dropbox ships experiences quickly and safely at global scale.

What tools support your day-to-day work?

Recently, Codex has been helping me a lot. It is connected to Jira, Confluence, Slack and GitHub, where most of the knowledge is stored for doing my work. I am able to use these tools to track execution, research and draft strategy documents, prepare presentation slides, create reports and the list goes on.

AI is used more and more within engineering as we attempt to innovate at scale. Dropbox team members are joining monthly sessions like Intro to AI Coding Tools and AI Security Essentials; attending AI Show & Tell; participating in hackathons and bootcamps; and using prior pilot formats, such as the three week Codex pilot with office hours and real-world experimentation.

How does your team experiment?

Experimentation starts with creating the time and space to experiment, having access to resources for experimentation, identifying interesting problems to solve and being comfortable with failure but using the learnings from there.

It is through experimentation that we built Nova — a coding agent platform which is now an essential part of our AI strategy.

This is all supported by Dropbox's broader approach to internal AI development through a program designed to increase AI fluency, with the practical goal of helping employees use AI in their everyday work, not just learn the theory. The program ecosystem combines self-paced learning, live training, tool-specific enablement, peer sharing and incentive-based application of skills.

How does your company adapt to change?

Adapting to change starts with understanding the reason for the change. Once that is understood, it is about creating buy-in and then giving people time to fully absorb the implications. A culture of trust is essential to adaptation and change and I have prioritized that consistently with my teams. A high-trust team can adapt faster than one that is not.

A specific example for me is how AI and agents have changed how work is done. It was important for engineers to understand why they needed to adopt AI, build trust in the tools, learn from peers who were already leveraging them, experiment with them for low-risk work and gradually incorporate them into their everyday workflows.

Our teams are already practicing this model through mini-hackathons and internal demos where teams build working tools and agents, then share them back with the broader team. Examples from the March 2026 mini hackathon included Nova integrated into the integrated development environment for background fixes and draft pull request generation, a Claude-based service scaffolding skill and automations that turn Jira tickets into Nova pull requests.

What’s it like to work on the AI and machine learning team at your company?

My recent AI work focuses on experimenting with running and operating LLMs on our own hardware. This goes beyond using AI to code, but involves understanding how the LLMs work and how to extract the best performance from the hardware for the nuances of each model.

Working on AI at Dropbox means solving real product problems, not building demos. Our team combines research with production engineering to create AI that helps people find information, make decisions and get work done more efficiently. At Dropbox, we’ve evolved from traditional retrieval systems to agentic AI that can reason, plan and take action across a user’s work. It’s a highly collaborative environment where ML engineers, infrastructure engineers and product teams iterate together, constantly balancing model capability, performance, quality and user trust as AI moves from experimentation into everyday workflows.

How is your team applying emerging technology in practical, business-relevant ways?

We have found that coding agents are useful for a wide range of tasks outside of coding. The sandboxing needed to provide safe autonomous coding agents translates directly into the necessary technology to provide more user-friendly agents for our customers.

One area we’re focused on is context engineering — ensuring AI has the right information at the right time to make better decisions. As Dropbox powered by Dash has become more capable, we’ve learned that simply giving models more context isn’t the answer. Instead, we consolidate retrieval, filter information aggressively and use specialized agents for complex tasks. These architectural decisions improve accuracy, speed and cost while enabling AI to search, summarize and act across the tools people already use every day. It’s a practical example of turning emerging AI techniques into reliable product experiences.

What should candidates know about the tools, collaboration or problem-solving involved in AI work at your company?

Engineering is constantly developing new agent skills and runtimes to help engineers get work done. What started as a skill to monitor your pull requests has turned into agents that address comments automatically. With each new shift in how we use AI, we integrate it more deeply into our development lifecycle. As we improve our own development flows internally, what we learn informs how we build our products.

Building production AI requires close collaboration across machine learning, backend infrastructure, search, product and design. We are considering user experience in all of our work, from the front end chrome down to the GPU. Engineers are encouraged to challenge assumptions, measure outcomes and share ideas openly. More importantly, you are encouraged to just try to build the thing you think will matter. The problems are technically deep — from designing agent architectures to optimizing context and performance — but they’re always grounded in delivering useful, trustworthy AI for customers.