Personal Vault Agent

Prepare a document list with a local agent. Review the choices. Decide what gets saved.

Comment

Preparing for a doctor's visit often starts with documents scattered across folders: a visit summary, a prescription, a referral. The immediate task is to decide which records belong together and keep that choice inspectable. Personal Vault Agent is designed for people doing that preparation on their own PC, with a local model and a visible decision before saving.

Our edited demo, under two minutes, follows one synthetic example: “everything about my migraine.” The agent proposes three fictional records. Vault Desk shows their names, its reasons and the keyword baseline. The operator accepts the set, confirms the intended audience, approves the exact preview and separately chooses Write. The saved manifest contains document references and hashes. The originals remain in the vault; nothing is attached or sent. VERIFIED means the written bytes match the approved preview, not that the document selection is complete.

The wider application also organises text documents into proposed cards with shelves, topics and issuers. A person can edit and confirm those cards, then ask for a document set in plain language. The agent proposes; the person reviews the choices and decides what gets saved.

We built two bounded agents with the Strands Agents SDK and an Ollama model running locally, qwen2.5:7b-instruct. Their tools list, read and propose. A separate Python kernel handles approval, a one-use 60-second permission window, writing and fresh read-back. The browser interface and terminal share the same operations. The agent cannot grant itself permission to write.

A local replay lets judges try recorded model responses without running a model. Sorting, migraine and tax are three independent recorded scenarios, not one continuous workflow. The model answers are recorded; the tools, operator choices, permission window, writer and receipt run now. A mismatch is visible instead of silently switching to a live model. The short film shows the migraine example and is not a new live-model evaluation.

The hardest work was keeping outcomes clear when the model or file operations misbehaved. Tool feedback had to help the model correct invalid fields. The interface had to distinguish permission from writing, and a verified file from a complete selection. In this example, keyword search finds the same three documents; that does not establish an advantage for the model. A separate recorded tax request missed receipts in both lists. That tax scenario is not shown in this film.

The owner directed this AI-assisted build; Codex and Claude produced code, tests and documentation. Codex operated the synthetic demo under the owner's direction. It is not an independent user test. There is no OCR, medical advice, demonstrated time saving or demonstrated benefit on real personal records.

The final film uses AI-generated narration and quiet piano. The delivered voice metadata identifies AI Voice Generator / AI Doc Maker, voice “delicate”; its underlying model is undisclosed.

Voiceover Generated with AIDOCMAKER.COM

At 00:48–01:06, an illustrative passage is visibly labelled “Edited close-up of matching contents - not a live scroll”. Its contents match the recorded list, but its source run is unidentified. This insert is not evidence of a continuous interaction from that same run.

Music credit

"Elementary" by Scott Buckley — released under CC-BY 4.0. Scott Buckley

Track: Elementary License: CC BY 4.0

Changes: the first 111.8 seconds were excerpted, converted to mono, faded at the beginning/end and reduced under narration. The music is synchronized to the Personal Vault demonstration, not distributed as a separate music release. No endorsement by Scott Buckley is implied.

Built With

  • ollama
  • python
  • strands-agents-sdk
Share this project:

Updates

Submission history