Confidant
A personal assistant that keeps private facts on the device. The cloud model writes with placeholders, Confidant fills in the real values locally, and every outbound call leaves a receipt.
- Status
- Status: Hackathon
- Stack
- Python, SQLite FTS5, Nemotron on Nebius Token Factory, MCP, Gradio
- Last commit
- 29 October 2026
- Try it
- Hosted demo on Hugging Face (replay)Video walkthrough, 2 min 40 s
- Sources
- Each tag opens the README line that makes the claim, at commit
2042c9b.
The problem
A useful assistant has to remember appointments, health, money and family. That is exactly the data you should not paste into a cloud model. README L18–19 (source in the Confidant repository) Any team that sends customer or staff data to a hosted model faces the same question: what left, and who can prove it?
What I built
- Memory with receipts: each memory records who said it, where it came from, when, and every correction. Answers cite memories. README L22–23 (source in the Confidant repository)
- A guard that classes memories as public, personal or private, and replaces private ones with placeholders before any call leaves the device. README L24–26 (source in the Confidant repository)
- Late binding: the cloud model never sees a private fact (proof: Confidant README L7–8). It answers with a placeholder, and Confidant fills in the real value locally.
- A receipt for every outbound call: provider, model, tokens, memories sent and withheld, and a SHA-256 of the exact messages. README L27–29 (source in the Confidant repository)
- An MCP server with 6 tools that never hands a client a private memory’s text. README L116–118 (source in the Confidant repository)
System as built
You ask
From the web UI, the CLI, or an agent over MCP.README L85 (source in the Confidant repository)
Routing on the device
Local rules first; a small model routes the rest, with values masked.README L69–70 (source in the Confidant repository)
Memory search
SQLite + FTS5, with provenance and versions.README L87 (source in the Confidant repository)
The guard decides
Public and personal memories pass; private ones become placeholders.README L88–89 (source in the Confidant repository)
The cloud model writes
Nemotron writes the answer; private items stay placeholders.README L89–91 (source in the Confidant repository)
Late binding
Confidant fills in the real values on the device.README L91 (source in the Confidant repository)
Receipt
Hash, tokens, and what was sent or withheld.README L93 (source in the Confidant repository)
How it is verified
- An offline test suite with all HTTP mocked. README L241 (source in the Confidant repository)
- Adversarial leak test: reviewers that did not write the guard wrote the attack rounds, and NVIDIA Nemotron 3 Ultra wrote round 9. Every finding became a test. All 199 attack messages now run in the suite and none leaks, per README. README L153–180 (source in the Confidant repository)
- A “Prove it” check rebuilds the last turn with private appointments moved and compares SHA-256 hashes. They are identical for every demo question. README L32–36 (source in the Confidant repository)
- The README calls this a regression result, not an estimate: no new attack round has been written against this version yet. README L180–182 (source in the Confidant repository)
What this means for your project
- If an assistant touches HR, health or finance data, a guard plus late binding keeps the values out of the hosted model’s context. That is an Agent or MCP build.
- Receipts turn “what did we send to the vendor today?” into a table, not a guess. README L30 (source in the Confidant repository)
- A Microsoft 365 and Entra security review covers the other half: what the connector itself can reach.
Status & limits
- Hackathon project, built for the Nebius × NVIDIA Global AI Hackathon, Personal AI track. README L14 (source in the Confidant repository)
- The hosted demo replays turns recorded live on 28 October 2026 and says so. With your own key, the same code runs live. README L10–11 (source in the Confidant repository)
- Synthetic persona only. The guard is a strong default, not a formal information-flow proof; paraphrases and relative dates are not detected. README L147–152 (source in the Confidant repository)
- Confidant writes its own receipts; they are not an external audit. README L227 (source in the Confidant repository)
- Last commit: 29 October 2026.
Links
- View the repo on GitHub
- Hosted demo on Hugging Face (replay)
- Video walkthrough, 2 min 40 s
- README at commit 2042c9b, as read for this page
Next case study: nexus-swarm-os
Want this pattern in your stack?
Contracted and paid on Upwork, at $90/h.