Back
Somni

Somni

Timeline

2025 - 2026, Stanford capstone

Role

Designer and builder

Team

Three

Scope

iOS app, bedside hardware, AI interpretation

Somni is a dream journal built for the people who need one most: those living with nightmare disorder and PTSD. I designed and built it across software and hardware, as my Stanford capstone.

Clinical UXHardware and software
  • Under 2 seconds from alarm to listening, against roughly 17 on a phone
  • Zero taps and zero typing, so the eyes never have to open
  • Eight enclosure versions to the final build, and 7 beta testers on the demo app

The film

The capstone film. Three minutes on the device, the app, and the eight enclosure versions behind them.

The product

Thirty seconds, eyes closed, and the reading is waiting in the morning

The device hears your alarm and starts listening. The app is where the pattern shows up.

The resolved form. Custom acoustic chamber, one amber LED.
The resolved form. Custom acoustic chamber, one amber LED.
Each dream comes back as an image you can revisit.
Home, the dream circle, and the gallery. Recorded off the working prototype, not mocked up.

Problem

The design problem is ninety seconds long

Storing dreams is trivial. The whole difficulty is that you are barely awake at the only moment the dream still exists.

Imagine

It is 3am and you have woken from the same nightmare again. The window in which you can still recall it is measured in minutes, and it closes while you are finding your phone, unlocking it, and hunting for the right screen.

For someone with nightmare disorder that scene repeats several nights a week. Clinicians treat these dreams with imagery rehearsal, which depends on the patient recalling the dream accurately. The recall is the treatment bottleneck.

50%of a dream is gone five minutes after waking, and 90% by ten

1. The material destroys itself while you fetch a tool

Nothing else in product design decays this fast. Every second spent reaching, unlocking, and navigating is taken directly out of the thing being recorded.

<1dream a week is what most people recall, if that

2. The baseline is so low that friction decides everything

At roughly one recalled dream a week there is no slack in the system. A capture flow that costs fifteen seconds does not reduce quality, it removes most of the data.

17sfrom alarm to the first word landing, on the phone path

3. By the time the phone is ready, the screen has woken you up

Reach, unlock, search, open, record. Five checkpoints, each one pulling the user further into the waking state that ends the recall window.

Recall decay after Hobson, and dream-recall frequency studies by Schredl, 2008 to 2011. Those numbers are why the design problem was never the journal. It was the ninety seconds after waking.

Users

The clinical case is the reason it exists. Three other groups were already building it themselves.

People treating nightmare disorder or PTSD have the sharpest need and the least tolerance for friction. Around them sit three groups who had already improvised a version, which is what told me the pattern layer was worth building.

New to the practice

The Curious

  • Already meditates or journals about waking life
  • Has wanted to dream-journal for years
  • The entry point is too steep at 3am

Already bought in

The Lapsed Journaler

  • Notebooks, voice memos, a Notes folder
  • Months of raw material sitting there
  • No way to see across any of it

Already bought in

The Hacker

  • Pastes transcripts into ChatGPT or Claude
  • Home-made prompts, spreadsheets, custom GPTs
  • Wants the analysis, not the duct tape

The third group is the clearest demand signal in the category. Free dream analysis with an LLM, here is my prompt is a recurring post on r/Dreams. People are hand-building this pipeline already.

The gap

Nothing on the market does three things at once

I audited everything in the category. Each product does one or two of these well and drops the third, and the third is always the one that matters at 3am.

Capture

  1. No typing, and nothing to press start on.
  2. No hunting through home screens for an app.
  3. It has to work in the sixty seconds before you are properly awake.

Insight

  1. A folder of audio files is not a dream practice.
  2. Themes, recurring symbols, and emotional patterns are the reason to return.
  3. One dream is noise. A month of them is a shape.

Screen free

  1. The bedside is sacred, and a screen is a portal to everything else.
  2. For a PTSD user the first thing you touch at 3am cannot be a feed.
  3. Light at 3am ends the recall window as surely as typing does.

Framing the solution

Mission

Design for the worst moment, not the demo. Everything else in the product is downstream of ninety seconds at 3am.

Three questions came out of that, and every decision below was argued against them rather than against a feature list.

How
might
we
How might we capture a dream from someone whose eyes are still closed?
How might we turn a pile of transcripts into something worth coming back to?
How might we put a device beside a bed without it behaving like a phone?

Approach

Thirty seconds, eyes closed, nothing to press

Every decision came off that 3am constraint. Typing is too slow and too demanding, so capture is voice first. The screen has to be readable by someone not fully awake, which set the contrast and the target sizes before I drew anything.

Alarm firesSomni wakesYou talkIt processesPatterns appear

  1. T + 0s

    Alarm fires

    The phone's own iOS alarm goes off, anywhere in the room. Nothing to pair.

  2. + 1.7s

    Somni wakes

    It recognizes the alarm pattern, lights a soft amber crescent, and prompts you to speak.

  3. + 3s

    You talk

    As soon as you are awake enough to speak, you describe the dream. Eyes still closed.

  4. + 30s

    The AI processes

    Transcript, themes, symbols and archetype tags, privately and on-device first.

  5. Later

    Patterns appear

    Open the app whenever. The reading has already been written for you.

Total user effort is about thirty seconds with your eyes closed. No tap, no typing, and deliberately no streak. A nightmare journal that punishes a missed night is the wrong object for this user.

Head to head

The difference is milliseconds, not minutes.

~17s

phone path, alarm to tapping record

<2s

Somni path, alarm to listening

The phone path is conscious and screen first

Five checkpoints before a single word lands, and by the end of them the screen has finished waking you up.

  1. + 3s

    Reach

    Find the phone. Dismiss the alarm.

  2. + 7s

    Unlock

    Eyes fully open now.

  3. + 11s

    Search

    Swipe through home screens to find an app.

  4. + 14s

    Open

    Voice Memos, Notes, or a dream app.

  5. + 17s

    Record

    The first word finally lands.

The Somni path is ambient and eyes closed

The device hears the alarm and starts listening. There is nothing to reach for, so there is nothing to fail at while half awake.

  1. < 2s

    Listening

    Somni recognizes the alarm tone and begins.

  2. Skipped

    Reach, unlock, search, open

    Four steps that only exist because a phone is a phone.

  3. < 2s

    Speaking

    You talk. Eyes still closed.

Fifteen seconds sounds like nothing until you price it against the decay curve. That window is the difference between a usable dream record and a fragment.

Hardware

A crescent on the nightstand, not a gadget

A phone is the wrong object at 3am. It is bright, it is a portal to everything else, and it demands a lock screen. So the product is a bedside device that hears your alarm, prompts once, records while the dream is still warm, and then disappears.

01

Hears your alarm

All 84 stock iOS alarm sounds, recognised in under two seconds. No pairing, no setup, no account to open before bed.

02

Records as you wake

One soft prompt. You speak the dream while it is still warm, with no screen involved.

03

Pushes the data

The recording goes to the app, where the transcript and the reading are waiting whenever you choose to look.

Enclosure iterationsThe working prototype inside it

The enclosure had to stop reading as a gadget

The constraint. Anything that looked like consumer electronics belonged on a desk, not beside a bed. It also had to hold a real acoustic chamber and an LED that confirms capture without lighting the room.

What shipped. A crescent. Soft enough to sit on a nightstand, deep enough to hold the microphone chamber, and lit by one amber ring that reads as attention rather than as a screen.

Version 5, printed and wired by hand. Three versions later the form resolved into the crescent above.

I taught myself the hardware for this, wiring and soldering included, and took the enclosure through eight versions to the final build. It is the part I would defend hardest, because it is where I found out what I do when nobody on the team already knows the answer.

Product

Five surfaces, and only one of them is the point

Home is the night's capture and the count that makes the practice visible. Circle surfaces the people who recur in your dreams, ranked by how often they turn up in them.

The surface that earns its keep

Insights reads a month back and returns the shape of it

Dominant themes, the emotional mix, recurring symbols, and the role you tend to play inside your own dreams. This is the layer the hackers were hand-building with pasted transcripts and custom prompts, and it is the reason anyone opens the app in daylight.

A single dream is noise. The pattern across months is the thing a user, or a clinician running imagery rehearsal, can actually act on.

Where it stands

It is not a concept deck. It is running.

In testing
7 beta users on the demo app, Stanford undergrads across mixed dorms
Hardware
Final enclosure built at version 8: functional crescent, custom acoustic chamber, amber LED
Software
A responsive app with a working front end and back end, deployed
Then
Polish onboarding, submit to the App Store, batch 10 to 15 devices

Somni was my Stanford capstone and it ended where capstones end: at the showcase, with the working device on the table and people who had never seen it trying to talk to it.

At the capstone showcase.
At the capstone showcase.

Reflection

Somni is the project I point at when someone asks whether I can build. Not prototype. Build: a real app with a database behind it, deployed, next to hardware I soldered myself and an enclosure I printed.

What it taught me:

Designing for a clinical population is mostly restraint. Every feature I wanted to add competed with the ninety seconds that actually mattered. The best decisions on this project were all subtractions, including the streak I deliberately never built.

The constraint chose the form factor. The device is not a device because hardware is interesting. It exists because a screen at 3am ends the thing the product is trying to capture, and no amount of app design gets around that.

The people improvising a version are the market. The clearest demand signal was not a survey. It was people pasting dream transcripts into ChatGPT with hand-written prompts, because that is what someone does when the product they want does not exist yet.

Next project

Killshot
All work