Most assistants answer questions about your day. Dunewind Wingmate carries it.
First seconds of the official clip today, the full clip Tuesday 20 October.
#DunewindAI#Wingmate#PersonalAI#ATES
Soon you won't open your personal assistant. You'll talk to it, cut in and joke with it. It'll speak first before your day slips.
Humans take turns in under a fifth of a second. That's the bar.
The next great PA app is maybe not your current IM app.
#DunewindAI#Wingmate#ATES
The productivity leak nobody budgets for is the other half of the day: the school strike, the jam before pickup, the storm warning at home.
A PA can carry all of it, if people trust it with their mail and family calendar. Here privacy decides whether it works at all.
@griswold@bot Your time-or-emotion test is right and I'd bet on emotion. Saved hours are nice. The real win is not carrying the week in your head. It starts when the meal plan already knows about Thursday's late meeting and the swim class.
@SoniaBaschez The 47 things live in places you'd have to open up: the inbox, the school emails, the family calendar, the bills. Nobody hands all of that to a product whose business is their data, so it only ever sees a slice. Trust is the bottleneck, not the model.
I asked Wingmate, our personal assistant, for a nearby store's hours: same answer as the big AI labs.
Then, the coupon in my emails. Different game entirely: Wingmate read my inbox privately, in a sovereign environment. Or you can use Vault AI, our private computing PC.
@sergeykarayev At $0.33/hour an always-on agent is ~$2,900 a year, so 1B of them is ~$2.9T at market rate. It's why we run the reading on a machine the household owns: a fixed cost, not a meter. And above all secure!
@Polymarket An agent with your inbox is the most valuable target on your phone. Patching helps, but the real fix is security by design, mainly where handling of private data happens: local models read the mail and the cloud only gets the facts a decision needs. This is better security.
@dhh Agreed. What remains expensive is knowing what an "achieved" job, per your criteria, looks like. Afterwards, the agent builds exactly what you wrote, including the vague parts. Privacy remains paramount.
@dps An agent that runs your calendar and inbox has to read everything to be useful. Most designs send all of it to the provider's cloud. We built our assistants the other way round: the mail is read by local models on infrastructure the household or company owns or sovereign clouds.
A year ago our CTO proposed a goal and I backed it: adapt the organisation so we could build a business app in under a month, with AI.
What we found: code was only 25–45% of our time. The rest: working out clients' needs, testing and deploying.
Coding faster isn't enough.
Anyone can claim AI builds software now. Almost nobody publishes the bill.
So that's what we're doing. One piece a week: what we built with Dunewind AI, what it cost in money and wall-clock time and what broke.
Started today. Including the parts that don't flatter us.
Every company I've built had the same bottleneck: what you need built is always 6 months and three hires away.
That's what Dunewind deletes. Not "AI writes code" — a product that is live and running at the end of the day.
I'll publish what it costs, including the failures.
I'm Alexandre, founder of ATES. We build Dunewind AI: you describe a product, it comes back live, hosted and running- in hours, not quarters.
This claim is easy to make and easy to check. So we're publishing the receipts: real costs, real wall-clock, real failures. Starting now.