A working model of how a person thinks, and software that defends their attention.
hello@headroomai.app
Headroom builds a working model of how you think and act — then defends the two things no agent can manufacture: your attention and your time.
Join the waitlistAn interruption is not free.
Interrupted work is not slower work. It is the same work done under more stress.
Re-entry after an interruption costs between three and thirteen seconds, and up to a minute when the task is demanding, measured in controlled laboratory tasks.1 Separately, interrupted work has been associated with faster completion and higher self-reported stress.2
A working day contains dozens of them. We have not measured either effect in real work; we are citing the literature, not our own data.
Then the tools arrive to help. Another inbox, another system to maintain, another place where commitments go to die.
Most of the cost lands on people already carrying more.
About 6.8% of adults show symptoms of ADHD and 2.5% meet full diagnostic persistence.3 Dyslexia affects roughly 7% of children and persists for many into adulthood, though we do not have one adult prevalence figure we trust.4 These are not rare conditions managed by a minority; they are ordinary colleagues.
They are the people whose cost is highest. They are not the only people paying.
The conditions overlap and are counted differently. The point is scale, not a sum.
The bill arrives as promises, not minutes.
A missed commitment costs trust. An avoided message costs a relationship or an income. Work never started costs the thing you were going to build.
Nobody itemises that bill, so it looks like a character problem. It is arithmetic on a system with fixed capacity, and the capacity changes daily. That is what we are solving: not making anyone more productive, stopping the loss.
A person, not a personality type.
Five layers, five conditions. Nothing here is a diagnosis, and nothing is a horoscope.
The layers describe how a mind works: state, identity, motivation, life load, cognition. The conditions describe what a life allows: material, social, bodily, temporal, moral. State outranks everything — what you can do today comes before what you should do today.
The five layers are our synthesis, not a finding. We assembled them from WHO's functioning framework, self-determination theory and the intervention literature. They are a hypothesis we test against outcomes, not a validated taxonomy, and we will change them when the data disagrees.
Three things, before anything else.
Not six agents and a plugin marketplace. Three jobs, done properly, on top of a model that makes them possible.
The model wants the whole life, not the inbox: calendar, mail, messages, the files you work in, the apps and sites that take the day, how you slept, where you were, what you spent. A model of a person that only sees their job is a model of their job.
What it may read is yours to decide, source by source, and revocable at any time. Breadth is the point: a system that watches one channel can only defend one channel.
Drafting works from mail you have already sent, with your explicit consent, so the reply sounds like you rather than like a model. It never sends as you without you, and it will not pretend to diagnose anything.
- 01Protects commitments
It knows what you promised, defends the time, and escalates before the promise breaks — not after.
- 02Starts the stuck thing
It turns an avoided task into the first physical action, then starts the clock with you instead of reminding you again.
- 03Drafts the avoided message
It writes the reply you have been putting off, in your own voice, ready for approval.
Tuesday, 09:03
One thing is on the screen: the message you have been avoiding since Thursday, drafted, with the sentence you would actually send.
Under it, one commitment it is defending at eleven, and a 25-minute block it has already moved twice. Nothing else: no backlog count, no streak, no red badge. It asks one question — enough for this today? — and if you say yes, it starts the timer with you.
You do not have to open it for it to work. The brief arrives each morning; the rest of the day it stays quiet unless a promise is about to break. There is no backlog to clear, because it never accumulates one: anything untouched for a week is dropped and said so, not carried as guilt.
“Hi Sam — sorry for the delay. Here is the version I actually want to send…”
A feature can be copied. A history cannot.
Two things changed at once. Models became able to act rather than suggest, and the platforms built those agents to optimise engagement — because engagement is what they sell. That leaves a gap for software whose success is measured by time returned rather than time spent, and whose incentives therefore point the other way.
The model spec is public in structure and expensive in content: every layer, condition and decision rule documented, versioned, and graded against the literature. A competitor can copy the shape of it. They cannot copy the work of getting each claim to survive review.
The second asset is the one that matters more: outcomes. What actually worked, for whom, in what state, on what day. A platform can ship a memory feature to a billion people. It cannot ship a consented history of what helped one person at 9am on a bad Tuesday — because that history has to be earned one approved action at a time, from someone who chose to share it.
That is the moat, and it is deliberately slow.
What we will not claim.
The category runs on assertions that do not survive replication. We grade every claim against the literature and refuse the ones the evidence does not carry, including the ones that sell best.
We can tell if you have ADHDNo validated adult classifier exists. We describe behaviour; we do not diagnose.Time blindnessFailed to replicate at n = 1,518. We build buffers and elapsed-time cues that work either way.Your social battery is at 20%No validated construct. We estimate load from self-report and contact volume, and show the uncertainty.21 days to a habitThe median is 59–66 days, range 4 to 335. We never promise a number.Personalisation works because it is personalAdaptive interventions average g ≈ 0.15. We test against a static version of our own product.
Built against a real problem, first.
Headroom is founded by Devangshu Mitra, who built products at HCL and studied applied economics and natural language processing at UT Dallas. The model on this page is the one he could not find anywhere else.
The research behind it is graded claim by claim; the refusals in §7 are the claims that failed. The model is written, the evidence is graded, and the first product is in build.
Join the first cohort
Or email hello@headroomai.app.
| Legal entity | Headroom AI Inc. |
|---|---|
| Incorporation | Delaware C corporation |
| Founded | 2026 |
| Founder | Devangshu Mitra |