Barge-in that works
Full-duplex audio in both directions. Start talking over Milo and it yields the floor instead of finishing its paragraph.
Personal AI desk companion
Say “Hey Milo” and stop typing.
Milo runs all day on the spare phone next to your monitor. It keeps your tasks, speaks your reminders out loud, reads and replies to your WhatsApp, knows where the family is, and roasts you when you dodge a lecture. The heavy work goes to Hermes, and code goes to Kiro.
- Full-duplex voice
- Hindi + English
- Hermes + Kiro
- Runs 24/7
12 Day StreakKiosk Active
--:--AM:--
Next reminderSubmit the report — 2:30 PM
Voice first
It's a conversation, not a command line.
Milo streams your microphone straight to a speech-to-speech model and streams the reply back as audio. You can interrupt mid-sentence and it stops talking — the way a person would.
Tools, not guesses
Milo doesn't pretend to save your task. It calls a real tool, writes to a real table, and your list updates while you're still talking.
Quiet when idle
Between conversations it's just a clock. Say the wake word, or tap Milo, and it comes back.
Tasks & reminders
“Remind me in two hours to submit the report.”
That’s the whole interaction. Milo stores the task with a future timestamp and schedules the alert. When the timer fires, the app wakes to the foreground, Milo appears, and it says the thing out loud.
Spoken alerts
Not a buzz you swipe away without reading. Milo pops to the foreground and tells you what's due.
Streaks worth keeping
Finish your Face Gym routine and the streak ticks up on the idle screen, where you'll actually see it.
Night mode
After hours the screen dims to deep plum, Milo curls up, and the room stops glowing. Try the toggle up top.
Bilingual by default
Your mom shouldn't have to speak English to it.
Milo understands Hindi and English and replies in whichever one you used. Tell it you’re stepping out and it switches into message-taking mode: anyone can talk to it while you’re gone, and it stores what they said with a timestamp.
When you get back, ask “any messages?” and Milo summarises them — in the language each person actually spoke.
“याद दिला देना कि उसे ये करना है।”
Milo: “Got it — मैं याद दिला दूँगा।”
Your phone, your people
Your WhatsApp and your family, one question away.
A companion app on your main phone feeds Milo your new WhatsApp messages. It reads them out, and when you tell it what to say, it replies for you. Ask “where’s everyone?” and you hear places, not coordinates.
Reads and replies
WhatsApp- WhatsApp · Rohanbhai DBMS notes bhej de
- MiloRohan wants the DBMS notes. Want me to reply?
- YouTell him I'll send them tonight.
- MiloSent: “Tonight pakka, bhej dunga.”
Where’s everyone?
Shared location- Bhaiya
- College4 min ago
- Maa
- Homejust now
- You
- At your desklive
Milo“Bhaiya’s at college and Maa’s at home. Everyone’s where they should be.”
Sarcasm, on the house
It notices when you're dodging the lecture.
Milo checks today’s screen time against the important things you keep pushing back. Postpone a lecture once and it lets it go. Postpone it while Instagram is at three hours and it stops being polite.
Today’s screen time
- 3h 04m
- YouTube
- 1h 12m
- 48m
PostponedDBMS lecture, 3 times
“Sir, you've spent 3 hours on Instagram today, and you're still single. The lecture you postponed three times is still waiting.”
Knows your screen time
The companion app sends Android's usage stats to Milo, so “how long was I on Insta?” gets a real answer, not a guess.
Tracks what you're dodging
Postpone a lecture or an important task again and Milo remembers. Overdue work doesn't quietly disappear from the list.
Roasts you, lovingly
It puts the two together. Celebrates your wins, calls out your excuses, and hypes you back into the lecture.
Heavy work, in the background
Hand it off. Hear the result.
Milo doesn’t make you wait on slow work. Email and research go to Hermes, code goes to Kiro, and you keep talking. When a job finishes, Milo tells you out loud.
Hermes
DoneGmail · research · job tracking
Reply to the Google email
Kiro
Ready for reviewCoding agent on your dev machine
Fix the landing page build
- Delegating doesn't block you: Milo returns right away and you keep talking.
- Ask “any update from Hermes?” or “what's Kiro doing?” whenever you like.
- Hermes reads your Gmail and sends the replies you ask for.
- Milo never says Kiro “shipped” something. Changes wait for your review.
Talk to Milo
Milo is speaking- YouAsk Hermes to reply to that email from Google.
- MiloSent to Hermes.
- YouAnd fix the landing page build.
- MiloSent to Kiro.
- MiloHermes replied to Google.
- MiloLanding build fixed. Ready for your review.
Always listening for “Hey Milo”Face Gym
A coach that notices you skipped a day.
Guided face and posture routines with Milo counting you through them. It’s the reason the streak exists — small daily things are the only ones that stick.
How it's built
Four pieces, one memory.
The desk phone is a thin client. All the intelligence (model streaming, tool dispatch, storage) sits in a relay you control, and the slow work runs on agents Milo can call.
The devices
A spare Android phone in kiosk mode runs the clock, tasks, Face Gym and a mic that streams over a WebSocket. A companion app on your main phone sends it WhatsApp, screen time and location.
The relay
A Node service that bridges audio to a speech-to-speech model, runs Milo's tools inline, and writes to your own tables. Self-hosted behind auto-TLS.
Hermes
A second agent on its own box for the slow work: Gmail, research, job tracking. It reads the same memory Milo does, so nothing gets re-explained.
Kiro
A coding agent on your dev machine. Milo sends it a brief, Kiro makes the change, and leaves it uncommitted so you review it before anything ships.
Coming next
Milo without the phone.
Because the relay holds the intelligence, the client can be almost anything. Next up is a dedicated ESP32-S3 device — a microphone, a speaker, a small screen, and no operating system to fight for always-on.
Same relay, same tools, same memory. Just a purpose-built body instead of a repurposed one.
Target build
- Board
- ESP32-S3, octal PSRAM
- Audio in
- I²S mic, 16 kHz mono
- Audio out
- I²S amp + speaker
- Transport
- Opus over secure WebSocket
- Display
- 320×240 clock face
Built for one desk. Mine.
Milo is a personal project, not a product launch — no signup, no waitlist. It’s here because a spare phone and a good model turned out to be a genuinely useful assistant.