Dropbox Tech Engineering Blog

Here's a concise summary of the blog post in 2-3 sentences: Dropbox's immutable blob store, Magic Pocket, faced a significant increase in storage overhead due to increased fragmentation and under-filled storage volumes after a new service rollout. The existing compaction strategy couldn't effectively reclaim space from the long tail of severely under-filled volumes, highlighting a limitation in the steady-state approach. Dropbox developed a multi-strategy approach to drive overhead back down, combining compaction, garbage collection, and efficient redundancy to achieve better storage efficiency and control at exabyte scale.

Dropbox's server monorepo grew to 87GB, slowing down engineering velocity due to long clone times and approaching the 100GB GitHub limit. The team discovered that Git's delta compression was inefficiently storing internationalization files, causing rapid repository growth. A locally tested fix using an experimental flag reduced the repository size to 20GB, but further investigation showed that this solution was incompatible with GitHub's server-side optimizations.

Dropbox's engineering team optimized Dash's relevance judge using DSPy, a framework for systematically optimizing prompts against a measurable objective. To adapt their existing judge for a lower-cost model, they defined a clear objective (minimizing disagreement with human relevance judgments while ensuring usable outputs) and used DSPy's GEPA optimizer to generate structured feedback for each example where the model disagreed with humans. This prompted a repeatable optimization loop, resulting in a more reliable and cheaper judge for production use.

Dropbox's Dash search engine uses a retrieval-augmented generation (RAG) pattern to generate responses, relying on large language models (LLMs) to analyze relevant content and ground responses. To improve search relevance, Dash pairs human labeling with LLM-assisted labeling, starting with a small amount of internal, human-labeled data and then amplifying efforts with LLMs to produce relevance labels at scale. This combination allows Dropbox to train Dash's search ranking models with high-quality labeled examples, resulting in improved search relevance and more accurate responses.

Dropbox engineers have adopted low-bit inference to improve the efficiency of their AI models, reducing memory and compute requirements. This technique involves quantizing tensors to lower precision, such as from 16-bit to 8-bit, which reduces memory footprint and enables faster processing, especially on NVIDIA GPUs with Tensor Cores. By leveraging these specialized cores, low-bit inference can double throughput and improve energy efficiency. Modern AI models, especially attention-based architectures, are computationally expensive due to repeated matrix multiplications. Low-bit inference helps mitigate this by scaling the performance of GPU cores, such as Tensor Cores, which can perform more operations per second at lower precision. This allows for efficient computation on large-scale linear algebra in neural networks.

Dropbox hosted an executive roundtable on AI and engineering productivity to share best practices with top companies. To accelerate progress, Dropbox tied AI tooling to tangible business results, prioritizing adoption and experimenting with various AI workflows. Results showed a significant increase in engineering productivity, with most developers using AI tools and a reduction in friction related to adoption.

Here's a 2-3 sentence summary of the engineering blog post: Dropbox developed Dash, an AI-powered platform that connects and analyzes content from various third-party apps and services within a single interface, using knowledge graphs and a highly secure data store. To build Dash, the team engineered a context engine that consists of connectors, a content understanding layer, and a graph-based model for storing and retrieving relevant information. By choosing index-based retrieval over federated retrieval, Dropbox was able to improve data freshness, access company-wide connectors, and enhance offline ranking experiments, but required significant custom work and handling of ingestion time challenges.

Dropbox Dash, a real-time AI-powered workspace, relies on a feature store to manage and deliver data signals to its ranking system. To meet its sub-100ms latency requirements and massive parallel read performance, Dropbox built a hybrid feature store using Feast, AWS DynamoDB, and Spark. The system serves features quickly, adapting to changing user behavior, and integrates with their existing infrastructure. The feature store architecture combines Feast's orchestration layer and serving APIs with a Go service for feature serving, cloud-based storage for offline indexing, and Spark jobs for feature ingestion and computation. This setup ensures a streamlined experience for engineers while abstracting away offline and online data management, pipeline orchestration, and data freshness guarantees. A three-part ingestion system was developed to balance freshness with reliability, allowing for real-time incorporation of new user signals while handling complex transformations and reducing infrastructure overload. The system consistently achieves p95 latencies in the 25-35ms range, making it possible to reliably meet Dash's latency targets

Here's a 2-3 sentence summary of the Dropbox 2025 summer intern class: This year's cohort of 43 interns at Dropbox's Camp Dropbox Intern Program made meaningful contributions to various teams, including engineering, AI, and search infrastructure. Through projects such as file history tracking, AI Sentinel, and cache optimization, interns tackled high-impact tasks aligned with company goals and cultivated growth, innovation, and connections. As part of the program, interns also benefited from over 6,000 hours of one-on-one mentorship, Virtual First events, and the Emerging Talent Summit, shaping their time at Dropbox and contributing to the development of Dropbox Dash, an AI-powered universal search product.

Here is a 2-3 sentence summary of the engineering blog post: To build a more intelligent and agentic AI, Dropbox's Dash uses context engineering to limit the information the model sees and focus on what's relevant. They achieved this through strategies such as providing a single interface for retrieval tools, filtering context to only useful information, and introducing specialized agents for complex tasks requiring deeper reasoning. As a result, Dash's performance has improved, and the model can make faster, better decisions, ultimately leading to better outcomes.