An ML file assistant: how Dropbox taught algorithms to tidy folders

Introduction

Picture this: you dump dozens of files into a shared folder, and the system suggests which subfolders they belong in — based on your habits and folder structure, not on file names. In November 2021 Dropbox shipped smart move: exactly that.

Under the hood: machine learning, human-centered design, and a pile of engineering trade-offs. We unpack how it works and what ML engineers can steal. Then we put the same skills to work: quizzes, pet projects, and vacancies inside IT Career Gym.

1. Why did smart move exist at all?

File chaos hurts — especially for admins and team leads. Moving files one by one is dozens of clicks, constant folder switching, and a chance to misplace something. Dropbox tried to automate the chore with one hard rule: the user stays in control.

The system suggests options in a clear UI; the person accepts, rejects, or edits. That is human-in-the-loop. It is ideal for experiments: you can test a hypothesis and watch real reactions without breaking the familiar flow.

The first version was also scoped tightly: move files from a chosen folder into its subfolders. That shrank the search space and simplified v1.

2. What is an “organized folder”? Spoiler: everyone means something else

The hardest stage is not the ML model — it is understanding users. File organization is deeply personal. In research, people said:

• “I do not trust others to touch my files”
• “If something is moved to the wrong place, I simply will not find it”

And the organization styles differ:

One model cannot cover every scenario. They had to narrow the focus and pick priority cases from research.

3. Privacy, finding data, and speed

• Privacy. Filenames are sensitive. Engineers cannot just look at real names. Training used only internal Dropbox data (with employee consent) and strict isolation protocols.
• Finding data. Suitable datasets had to be filtered by hand, dropping auto-generated app folders that did not match the scenario.
• Speed. Suggestions are served synchronously on the user request. Any delay hurts the experience, so latency work was critical.

The same constraints show up in AI-engineer interviews: feature privacy, synthetic labels, and inference right now. See 45+ questions for AI engineers.

4. Why did the old model fail?

Dropbox already had a prototype — suggested destinations — that predicted a folder for a single file from recent activity. They tried it for smart move. It failed.

Problems:

• Instability — suggestions changed depending on where the user had just browsed. Predictability is a trust factor.
• Opacity — people did not see why this folder was chosen. When they organize, they look at file and folder names first, not navigation history.
• No priorities — testers needed a confidence sort: “high” shown immediately, “medium” second, “low” hidden.

Takeaway: do not reuse a model built for a different job, even if the tasks look similar. A prototype is a cheap way to learn that.

5. Building the new model: data first

The core question: where do you get training data if you are predicting something that has not happened yet?

The elegant move: find folders that were already organized in Dropbox and treat them as the “ideal end state” after a successful smart move.

Example structure:

root
├── folder_A/
│   └── file_1.pdf
├── folder_B/
│   └── file_2.pdf
└── folder_C/
    ├── file_3.jpg
    ├── folder_D/
    └── folder_E/
        └── file_4.pdf

From that tree they generated training pairs: (file, candidate folder, label 1 = correct, 0 = not). That produced millions of examples from internal data: Dropbox employees actively use the product.

Which signals did they use?

They looked for a minimal feature set that can be computed on the fly:

• The filename being moved (with extension)
• Each candidate folder name
• File and folder names inside each candidate — “neighbors” give context (if the folder is already full of images, a new .png belongs there)

Neighbors often beat the folder name itself. A folder can be called Finance and still contain w2.pdf, taxreturn_2020.pdf, charitable.jpg. The model learns that inner atmosphere.

6. Model architecture (simplified)

1. Tokenization — split file and folder names into tokens.
2. Embeddings — a custom encoder on character and GloVe embeddings. It captures semantics (tax and receipt are close) and file types (.png, .jpg, .img are all images).
3. Similarity — comparison matrices for (file → folder) and (file → neighbors).
4. Features — basic engineering, sometimes with penalties for certain types.
5. Deep net — fewer than 20 layers with dropout, trained on internal data.

The output is a score for each (file, folder) pair. Rank, take top-1 as the suggestion. Then cut by confidence: top 20% is “high”, the next band is “medium”, the rest are hidden.

If you build a similar ranker for a portfolio, publish it as a case in the pet-project catalog — recruiters care more about a confidence threshold and UX than “yet another net”.

7. Versus a heuristic: who won?

They compared two approaches: a simple name-similarity heuristic and the neural net above.

Offline evaluation (internal data):

Here the model wins. In an online test with real users the picture flipped:

Why did the heuristic suddenly beat the model? Most likely they overfit Dropbox-internal data, while real users organize files differently. They did not despair: the same model later worked better in other prototypes (for example the Save-to-Dropbox browser extension).

8. What next — and what if you are building something similar

Two main conclusions:

1. ML assistants for organization are promising only if the user stays in control. Do not replace the person — help them.
2. Prototyping is your best friend. Put the model in real users’ hands as early as you can to surface non-obvious problems.

Dropbox’s later plans: alternative UX patterns that do not overload the UI, and fine-tuning an LLM against the current model. In 2021 that was harder; now it is realistic — especially for non-English filenames.

Takeaways for ML engineers

Want to lock this in with practice? On IT Career Gym the same topics live in quizzes, pet projects, and vacancy matches — from experiment to monitoring, as in the MLOps article.

FAQ

Conclusion

Smart move is not “AI magic”. It is a narrow scenario, synthetic labels, fast inference, and an honest online test. An ML assistant works when the user stays in the loop.

If you are building something similar, start with a prototype, a confidence threshold, and an acceptance metric — not only accuracy. Those same questions show up in interviews and ML-engineer vacancies.