work / pip · Live · daily use
Pip
A JARVIS-style assistant that actually runs my day.
Play with it
A pure-frontend replica with mock data: nothing is saved, nothing leaves the page. The real app is richer; this is the feel of it.
loading the demo…
The problem
Calendar, tasks, habits, goals and notes scattered across apps, and admin that drains the energy real work needs.
What it does
A chat and real-time voice assistant sharing one brain and one memory. It reads and writes Google Calendar and Tasks, tracks habits with streaks, links them to goals, takes notes, auto-schedules work around no-go zones, and stages every AI write behind an approval queue. The interface is a command-deck HUD with an audio-reactive voice orb.
Under the hood
- One brain for chat and voice: a single tool loop over a single memory, so the two channels never disagree
- The LLM is a config seam: the model was swapped four times (Gemini, Kimi, Groq Llama, back to Gemini) with roughly a one-file change each time
- Voice time-to-first-word cut from about five seconds to under two through streaming, batched reads and warm-up calls
- An approval workflow stages calendar, task and note writes for review, so the assistant earns trust instead of assuming it
The thinking
Target user
Me first: a founder with an ADHD-I brain juggling two businesses, study and eight builds. By extension, anyone whose working memory needs an external system that talks back.
Pain points
- Calendar, tasks, habits, goals and notes live in five apps that never talk to each other
- Typing and admin drain the exact energy the real work needs
- Assistants forget: what you told the chat, the voice call never heard
Gap in the market
Consumer assistants are either chat-only or voice-only, and their memories are split. Nothing affordable runs one brain across both channels with real write access to your life, gated by trust controls.
AI + tooling
- Gemini 3 Flash
- ElevenLabs realtime voice
- Claude Code (build)
- Supabase Postgres
- Next.js
- Web Push
Frameworks
- Jobs-to-be-done: an external working memory that talks back
- Fogg behaviour model: nudges land at the moment of ability, not just intent
- Habit loop design (cue, routine, reward) with streaks linked to goals
- Progressive trust: staged approvals before the AI writes to your life
- One-brain architecture (single tool loop, single memory) + idempotent tools
Success metrics
- Daily driver since launch
- Voice time-to-first-word under 2s (from ~5s)
- Model swapped 4x with ~one-file changes
- TBD: weekly actions executed via Pip
A single header prefix cost weeks, disguised as random voice disconnects. Distrust the symptom, read the wire.