Back office ops · Production

Drew Houston on 400 hours coding with LLMs and Dropbox's AI-first transformation

The problem

Knowledge workers face growing information overload with files scattered across multiple apps and screens, making retrieval difficult; earlier classical NLP and tiny pre-GPT LLMs broke down on free-form human language, forcing Drew to abandon multiple automation projects.

First attempt

Classical NLP fell apart on natural language, and very small pre-GPT models hallucinated and repeated themselves, causing Drew to write off the approach before the GPT-3/4 era made reliable text understanding possible.

Workflow diagram · grounded in source
1
File preview triggers AI Q&A
trigger
“when you preview a file on Dropbox, you can have kind of the most straightforward possible integration of AI, which is a good thing. Like basically you have a long PDF, you want to be able to ask questions of it”
2
RAG answers file questions
ai_action
“a pretty basic implementation of RAG and being able to do that when you preview a file on Dropbox”
3
Dash aggregates all apps
integration
“You can connect Dash to all your applications as well as other storage providers like Google Drive, OneDrive, etc without having to move your files over”
4
Semantic search surfaces results
ai_action
“you don't even have to get the keywords right. The relevance and ranking is super good. We even like untuned”
5
Request proxied to LLM router
routing
“I use the continue.dev just like AI chat UI basically as just the UI layer, but I also proxy the request. I proxy the request to my own backend, which is sort of like a router. You can use any backend.”
6
Files injected as LLM context
ai_action
“there's a separate thing where I can say like include all these files by default with the request. And then it becomes a lot easier and like without like cutting and pasting”
Reported outcome

Drew built a working personal semantic search engine within days and shipped Dropbox AI (File GPT) with RAG in 2023; relevance and ranking were described as super good even untuned, and Dropbox Dash became a universal search product connecting all apps without file migration.

Reported metrics
CEO hours coding with LLMsover 400 hours
Reported stack
VS CodeCursorcontinue.devSonnet 3.5LlamaVLLMSGLangXLlamaNext.jsFlaskSQLiteDropbox DashRAGGPT-4Google DriveOneDrive
Source
https://www.latent.space/p/drew-houston
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Drew built a working personal semantic search engine within days and shipped Dropbox AI (File GPT) with RAG in 2023; relevance and ranking were described as super good even untuned, and Dropbox Dash became a universal…

What tools did this team use?

VS Code, Cursor, continue.dev, Sonnet 3.5, Llama, VLLM, SGLang, XLlama, Next.js, Flask.

What results were reported?

CEO hours coding with LLMs: over 400 hours (source-reported, not independently verified).

What failed first in this deployment?

Classical NLP fell apart on natural language, and very small pre-GPT models hallucinated and repeated themselves, causing Drew to write off the approach before the GPT-3/4 era made reliable text understanding possible.

How is this back office ops AI workflow structured?

File preview triggers AI Q&A → RAG answers file questions → Dash aggregates all apps → Semantic search surfaces results → Request proxied to LLM router → Files injected as LLM context.

Related back office ops cases