Back office ops · Production

Dropbox machine intelligence initiative (DBXi): OCR, personalized search, and intelligent workspace

The problem

Knowledge workers face information overload from content scattered across many cloud services, spending too much time on organization, contextualization, and prioritization rather than focused work.

Workflow diagram · grounded in source
1
Connect content sources
trigger
“files but also content like Google Docs, as well as collaborative activity in emails, messaging apps, and calendars—whatever a user chooses to connect to Dropbox”
2
OCR document scanning
ai_action
“custom optical character recognition (OCR) pipeline to help our users quickly scan and find their content. We combined classical machine vision techniques with advanced deep learning methods to create a mobile scanner experience that was…”
3
ML-powered personalized search
ai_action
“completely rebuilt our search infrastructure to improve the quality and speed of results from the hundreds of billions of pieces of content our users entrust to our platform. Because of our granular sharing permissions, each user has a u…”
4
User-specific graph traversal
ai_action
“traversing a user-specific graph that connects people, content, and activity signals in privacy-preserving ways”
5
Prioritize and cluster content
ai_action
“we cluster content and collaborative activity across silos so users can immediately see which projects need their attention and be only a click away from the content they need”
6
Intelligent activity feed
output
“intelligently highlights their most important work connected to Dropbox”
Reported outcome

Dropbox built a custom OCR pipeline faster and more accurate than off-the-shelf solutions, rebuilt its search infrastructure for improved quality and speed, and developed a DBXi prototype that clusters and prioritizes work content in an intelligent activity feed.

Reported metrics
OCR accuracy vs off-the-shelffaster and more accurate than any off-the-shelf solutions we could find
Reported stack
OCRmachine visiondeep learningGoogle Docs
Source
https://dropbox.tech/machine-learning/machine-intelligence-at-dropbox-an-update-from-our-dbxi-team
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Dropbox built a custom OCR pipeline faster and more accurate than off-the-shelf solutions, rebuilt its search infrastructure for improved quality and speed, and developed a DBXi prototype that clusters and prioritizes…

What tools did this team use?

OCR, machine vision, deep learning, Google Docs.

What results were reported?

OCR accuracy vs off-the-shelf: faster and more accurate than any off-the-shelf solutions we could find (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Connect content sources → OCR document scanning → ML-powered personalized search → User-specific graph traversal → Prioritize and cluster content → Intelligent activity feed.

Related back office ops cases