Open to Summer 2027 · data / analytics / AI engineeringList view

Vesper — Personal AI assistant

Daily use×

An always-on AI assistant that knows my notes, deadlines and inbox, works from every device I own, and logs every action it takes.

23,744 events
logged by daily use since Sep 2, 2026
  • 39MCP tools
  • 321automated tests
  • 46architecture decision records
  • 5clients: desktop, iPhone/iPad, terminal, Telegram, Claude Code
Vesper's Intuition pane: review-mode delivery with 1,723 nudges raised, 380 judged and 32% judged useful, a confidence histogram with a draggable review floor at 0.75, and loop settings for interval, cooldown and minimum confidence.
The intuition engine, graded: every nudge arrives with verdict buttons, and the threshold is tuned on the judged history.
Vesper's Log pane filtered to tool events: numbered tool.call and tool.result rows for WebSearch, Read, Write and Glob, each with a status and duration.
The append-only event log every other view is built from, here filtered to tool calls (failures included).

The problem

Siri knows nothing about my life and can't act on it. I wanted one assistant that reads my notes vault, school deadlines and email, works the same from my laptop, phone and terminal, points out what I'm forgetting — and keeps working when one AI provider hits its limit.

What I built

  • One headless brain, thin clients. A Fastify daemon (about 80 HTTP/WebSocket routes) serves a Tauri desktop app, a terminal CLI, a native SwiftUI iPhone/iPad app with widgets and a Share extension, and a Telegram bot. Every action is written to an append-only SQLite event log before anyone is notified.
  • Tools and model routing. 39 MCP tools (vault search, semantic recall, reminders, scheduling, screen vision). A router sends each request to Claude, Codex or a local Qwen model and fails over when a quota runs out — 44 of 108 routed turns were answered on-device.
  • An intuition engine that gets graded. A local model checks deadlines and the vault on every tick and raises nudges I can mark useful, wrong or noise in one tap: 1,725 nudges raised and 343 judged by hand, a dataset that drives the next fix (the nudge study).
  • Local-first memory and senses. Hybrid semantic + keyword search (local embeddings, FTS5, reciprocal-rank fusion) over 5,065 chunks, fed by voice captures routed into a ~2,000-note vault that Vesper searches; on-device speech-to-text (~1.2 s for a 2.5 s clip) and screen OCR. None of it leaves the Mac.
  • Built to last. 46 ADRs, 321 tests, secret scanning and dependency audits in CI. A security audit fixed 14 advisories, and a watchdog cut detection of a stuck local model from 12 days to about 75 seconds.

How it works

  1. Desktop · phone · terminal · Telegram
  2. Daemon (auth, routing)
  3. Claude / Codex / local Qwen
  4. 39 MCP tools
  5. Append-only event log
  6. Fan-out to every client

Stack

  • TypeScript
  • Node.js
  • Fastify
  • SQLite + FTS5
  • MCP
  • Claude Agent SDK
  • Ollama (Qwen3)
  • Tauri + Rust
  • React
  • SwiftUI
  • WidgetKit
  • whisper.cpp
  • Python
  • GitHub Actions

Shows

Proof

Real usage23,744 events over 201 sessions, Sep 2 – Oct 14 2026aggregate count from the event log
Human feedback1,725 nudges raised · 343 judged (120 useful / 35 wrong / 188 noise)event log, Oct 15 2026
Model routing108 routed turns: 64 to Claude, 44 to a local modelevent log
Engineering91 commits · 46 ADRs · 321 tests (211 Node, 110 Python)git history and test files
Reliabilitystuck-model detection 12 days → ~75 sADR 0042
Security14 advisories fixed (5 high, 1 critical); every route but /health requires a tokensecurity audit log, Oct 12 2026

What I'd do next

Move the brain to an always-on host and give the intuition engine a proper evaluation set built from the 343 judged nudges.

Sep 2026 – present · Private repo — it runs on my own data. Screenshots and recordings only; no code links.

← Back to the desktop All projects