You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$3,900 left to reach the funding goal.
Two clients can both have an “Invoices” folder. A document may clearly be an invoice, but that still doesn’t tell you whose it is. Sometimes the useful answer is “I can’t tell — please choose.”
I’ve been working on this inside Autofilot, a Mac app that helps documents find their place in the folders you already use. I’m asking for $3,900 to make its local folder-selection logic available as an open-source library and command-line tool.
This came out of my own agency work. On 4 November I asked why nine files still needed manual choices. After two defects were fixed, a controlled check on copies put six readable documents in the right folders without manual choices. Three audio files with insufficient information still needed a person to choose. That distinction matters to me: useful automation also needs to know when to stop.
Another developer could give it document text, possible destinations and examples of documents already filed there. They’d get a suggested folder, or a request for review, with the reasons. They could inspect and change the logic without buying Autofilot or uploading documents to a hosted service.
This 46-second demo shows automatic filing, history and Undo in the current app, using synthetic documents:
https://autofilot-native-demo.nb-649d.chatgpt.site
The separate public package is what this grant would make possible.
The first use I have in mind is a developer adding folder suggestions to a local Mac tool, or someone connecting the command-line tool to their own filing script. They could try a decision before letting their software act on it.
Local filing tools already exist. For example, organize supports content-based rules and simulated actions. I want to make Autofilot’s context-and-example-based decisions separately available; I haven’t established that they work better than other tools.
The release would include a Swift library, a read-only command-line tool, documented inputs and outputs, and runnable examples of a clear choice and an ambiguous one. It would take already-extracted text, make no file moves and need no account, backend or API key.
Week one would cover the public interface, cases and initial extraction. Week two: implementation, integration and error cases. Week three: independent review, fixes, documentation and release.
I wrote the code and can release the selected component under MIT. The public release would include the library, command-line tool, examples, tests and instructions. The initial target is macOS.
An independent reviewer would build the package from the instructions, run the examples and add a new synthetic case without the private app or help from me.
We’d publish expected and actual outcomes for 36 authored cases, including 12 held back from tuning. These would cover clear matches, ambiguous ownership, conflicting examples, technical folders, missing text and changed context. Correct choices, appropriate refusals and unsafe suggestions would be counted separately. A predefined safety failure would block release until fixed.
The reviewer would also check operation with network access denied and that default output avoids raw document text and private absolute paths. This tests whether the package is independently usable; wider demand and real-world accuracy remain open questions.
The minimum and full goal are both $3,900, covering one complete release:
• $500: two days on the public interface, dependency boundary and test cases.
• $1,500: six days extracting the library and building the command-line tool.
• $500: two days on the integration example, compatibility checks and fixes.
• $750: three days on build instructions, review fixes, examples and release.
• $600: two days of independent review at $300.
• Up to $50: transfer and currency-conversion costs.
My time is 13 days at $250, gross before applicable personal tax. The reviewer isn’t booked yet; their rate is an estimate. I already have the equipment and tools.
The commercial app could reuse the public release on the same licence as anyone else. I’d start after approval, the grant agreement and sufficient cleared funds, with no planned personal cash contribution. If required costs exceed the budget, I’d agree changes before accepting funding.
I’m Yuri Negrebetsky, based in London. I write music and build small useful apps. I designed and implemented Autofilot using AI-assisted development.
I want these tools to help inside what a person is already doing. You carry on working and the files find their place. Ideally, you mostly forget the app is there.
I’d do the extraction, testing and documentation. An independent reviewer, not yet appointed, would check the public package.
The code could be harder to separate than expected. If it needs a much larger rewrite, I’d pause and agree a revised scope with Manifund.
Nobody outside the project has committed to adopting the component yet. The reviewer can test usability, but that won’t establish wider demand. I’d report this limit and failures in the tests.
The current Autofilot app is closed-source and provided by NAURAL PROJECT LTD. This grant would open the local component under MIT. The app’s optional cloud-analysis path is outside this release. Unspent funds would be handled under the grant agreement.
Autofilot has received no external funding in the last 12 months.
On 10 November 2026 I sent FUTO a $3,900 proposal for this same stage. I also submitted a Mozilla Fellowship letter of intent for a broader Autofilot project. Neither has resulted in an award as of this draft.
If FUTO funds this stage, I’d close or revise this request before accepting overlapping support. Completed work would become the starting point for a later Mozilla project, with its scope and costs updated.