JUN 1, 2026
EngBrief
Search⌘K
LatestTopicsSourcesSaved
Eng&Brief

Engineering insights from the world's best tech companies, curated and summarized.

Weekly brief

Browse

TopicsSourcesFavorites

More

SearchRSS Feed
© 2026 EngBriefUpdated every 4 hours
← Sources
dropbox.tech icon
Distributed Storage

Dropbox Tech Blog

12 articles on EngBrief

Dropbox Tech explores how the company builds and scales its file sync, collaboration, and storage platform. Posts cover distributed storage systems, desktop client engineering, machine learning for content intelligence, security architecture, and the migration from AWS to their own custom infrastructure.

Distributed StorageDesktop EngineeringMachine LearningInfrastructure
Visit blog →

Latest Articles

Dropbox3d ago

Beyond code generation: rethinking engineering productivity in the age of AI agents

Dropbox has moved beyond solely using AI coding tools to accelerate code generation and is now focusing on creating agentic systems that can execute scoped tasks. This shift has changed the way engineers work, with implementation workflows becoming more parallel and repetitive execution being offloaded to AI agents. As a result, Dropbox has built a platform called Nova, which allows engineers to describe tasks in plain language and run AI agents in a controlled environment, producing meaningful output and changing the operating model of software development.

InfrastructureScale
1 min
Dropbox10d ago

Introducing Nova, our internal platform for coding agents

Dropbox built Nova, a platform for coding agents to assist engineers in various workflows, including development, remediation, and migration. Nova enables developers to run multiple coding sessions in parallel, supporting both interactive and background jobs, while maintaining consistent execution, validation, and context handling. The platform integrates with internal tools and workflows, such as Bazel and Slack, and provides features like prompt evaluation, observability, and feedback collection to improve agent performance.

InfrastructureScale
1 min
DropboxApr 2, 2026

Improving storage efficiency in Magic Pocket, our immutable blob store

Here is a 2-3 sentence summary of the blog post: Dropbox's immutable blob store, Magic Pocket, experienced a significant increase in storage overhead due to a new service causing under-filled volumes, which were not efficiently reclaimed by the existing compaction strategy. To address this, Dropbox developed a multi-strategy approach to drive overhead back down below the previous baseline, combining a new compaction strategy for dealing with under-filled volumes with existing garbage collection and compaction processes. The new solution enables faster and more efficient reclamation of space, crucial for maintaining cost-effectiveness and scalability in an exabyte-scale storage system.

InfrastructureScale
1 min
DropboxMar 25, 2026

Reducing our monorepo size to improve developer velocity

Dropbox reduced their monorepo size by 77% from 87GB to 20GB, cutting the time to clone the repository from over an hour to under 15 minutes. The growth was caused by Git's delta compression, which was incorrectly pairing similar files due to a directory structure mismatch with their internationalization (i18n) files. Dropbox worked with GitHub to implement a solution that used a more aggressive repack with tuned window and depth parameters, resulting in a 84GB repository being compressed to 20GB.

InfrastructureScale
1 min
DropboxMar 17, 2026

How we optimized Dash's relevance judge with DSPy

Here's a 3-sentence summary of the blog post: Dropbox engineers optimized Dash's relevance judge using DSPy, an open-source framework for systematically optimizing prompts against a measurable objective, to adapt the judge to a lower-cost model while minimizing disagreement with human relevance judgments. They defined a clear objective, minimized NMSE and formatted failures, and used DSPy's GEPA optimizer to generate feedback and iteratively improve prompts for the new model, gpt-oss-120b. This approach enabled them to maintain performance, reduce cost, and improve reliability in production, ultimately achieving better alignment between model outputs and human ratings.

InfrastructureScale
1 min
DropboxFeb 26, 2026

Using LLMs to amplify human labeling and improve Dash search relevance

Dropbox's Dash uses a retrieval-augmented generation pattern to provide accurate responses, relying on large language models to analyze search results. However, generating high-quality relevance labels for training is a challenge, especially since there are millions of documents in enterprise search indexes. To address this, Dropbox amplifies human labeling efforts with LLMs, leveraging their ability to produce relevance judgments at scale.

InfrastructureScale
1 min
DropboxFeb 12, 2026

How low-bit inference enables efficient AI

Dropbox is leveraging low-bit inference techniques to improve the efficiency of their AI models. By reducing the precision of numerical values in their attention-based architectures, they can significantly lower memory usage and energy consumption while maintaining model performance. This approach, combined with optimized hardware utilization, allows Dropbox to deliver fast and cost-effective AI-powered search and understanding across vast amounts of user content.

InfrastructureScale
1 min
DropboxFeb 11, 2026

Insights from our executive roundtable on AI and engineering productivity

Dropbox's executive roundtable on AI and engineering productivity addressed key challenges in effectively using AI in workflows. To achieve tangible business results, Dropbox prioritized AI adoption company-wide, empowering teams to experiment with tooling and reducing approval overhead. This effort led to improved productivity across the software development life cycle, with most developers now using at least one AI tool, and increased PR throughput per month per engineer. Effective use of AI requires a balanced approach to productivity gains, quality, and maintenance costs. Leaders play a crucial role in enforcing AI usage norms and formalizing AI competency within career frameworks. As a result, Dropbox has seen a significant increase in positive sentiment among engineers and plans to focus on mapping productivity gains to tangible business outcomes in 2026.

InfrastructureScale
1 min
DropboxJan 28, 2026

Engineering VP Josh Clemm on how we use knowledge graphs, MCP, and DSPy in Dash

Dropbox engineers built Dash, a knowledge repository that integrates content from various third-party apps and Dropbox, utilizing a knowledge graph engine. To achieve this, they created a context engine that includes connectors for extracting data, a content understanding component for enriching and normalizing content, and a knowledge graph builder for modeling this information as a graph. This allows Dash to provide personalized and ACL'd results for users. They chose index-based retrieval over federated retrieval to gain access to company-wide connectors and pre-process content for enriched data sets. However, this approach requires significant custom work, ingestion time freshness issues, and storage costs. To overcome challenges with MCP at Dropbox scale, they limited tool definitions to 100,000 tokens to prevent "context rot" and degradation in agent effectiveness. They also compared MCP implementation with raw index retrieval and found the raw index to be much faster, with results coming back within seconds compared to MCP's 45 seconds.

InfrastructureScale
1 min
DropboxDec 18, 2025

Inside the feature store powering real-time AI in Dropbox Dash

Dropbox built a feature store to power real-time AI in Dropbox Dash, which relies on a ranking system powered by machine learning to find relevant files, chats, and conversations. The feature store serves data signals, or "features," to models, and is built to handle massive parallel reads and strict latency budgets. To achieve this, Dropbox designed a hybrid feature store using Feast, which leverages Dynovault for low-latency feature lookups and Spark jobs for feature ingestion and computation. The feature store architecture consists of three components: Feast for orchestration and serving APIs, a custom Go service for feature serving, and Dynovault for cloud-based storage and feature lookups. This setup enables true concurrency and meets Dash's sub-100ms latency requirements. To keep features fresh, Dropbox built a three-part ingestion system that balances freshness with reliability, using batch ingestion for complex transformations and real-time detection for modified records.

InfrastructureScale
1 min
DropboxNov 26, 2025

Building the future: highlights from Dropbox’s 2025 summer intern class

Here's a 2-3 sentence summary of the Dropbox 2025 summer intern class: Dropbox's 2025 summer intern class, comprising 43 students from 27 colleges and universities, participated in the company's Camp Dropbox Intern Program, where they gained over 6,000 hours of one-on-one mentorship and worked on high-impact projects aligned with team and company goals. Interns contributed to the development of the AI-powered universal search product, Dropbox Dash, and made meaningful technical wins, such as streamlining metadata infrastructure and reducing operational costs, improving front-end latency, and automating code migrations. The intern program is thoughtfully designed to cultivate growth, spark innovation, and build lasting connections within the engineering organization.

InfrastructureScale
1 min
DropboxNov 17, 2025

How Dash uses context engineering for smarter AI

Here's a 3-sentence summary of how Dash uses context engineering for smarter AI: By applying context engineering, Dash transformed from a traditional search system to an agentic AI that can reason and act effectively, requiring structured, filtered, and delivered context at the right time. To achieve better outcomes, Dash employs three core strategies: limiting tool definitions, filtering context to only what's relevant, and introducing specialized agents for complex tasks, ensuring the model receives the right information at the right time in the right form. This approach enables Dash to make faster, better decisions and provides a consistent, efficient, and scalable way to manage context for agentic AI systems.

InfrastructureScale
1 min