Parallel

Your attention digital twin

Every ping forks your day. Parallel shows you both futures before you tap.

Built for CASCADE ’26 · Theme 3 · Problem 3: The Attention Digital Twin

Run it · All features · Project context · Logic & formulas · Psychology · Phone experience · Door to the Future · Real notifications · Architecture · Data & ML · Mobile build · API

---

Contents

1. The problem in one example
2. What Parallel is
3. See it on a phone in 15 seconds
4. Feature tour
5. The phone-first, Gen Z layer
6. How it would run over real notifications
7. Data & ML (honest version)
8. Quick start
9. Architecture
10. Repository map
11. What is built, what is simulated, what is next
12. Documentation index

---

The problem in one example

You're studying. “aarav_07 and 12 others liked your post.”

Universe A: you tap it	Universe B: your twin holds it
+2 min to look → Instagram opened → +17 min of reels → +8 min to remember where you were = ~27 focus minutes gone	Held until your next checkpoint, read in a 10-second batch = ~0.7 min

Everyone's attention is different. The same like can cost one person 2 minutes and another 40. Parallel learns you: it predicts the chain of events one notification will set off, gives a range and a confidence, compares it with the counterfactual (what happens if you hold it), and then checks itself against what really happened.

What Parallel is

Parallel is a digital twin of your attention. It is three things stacked on top of each other:

Layer	What it does	Where it lives
The model	Learns when you focus best, what pulls you away, which apps fragment you, when you're tired and which pings actually matter. Simulates 2,000 futures per ping with Monte Carlo, and estimates the counterfactual from randomized holdouts.	backend/app/twin.py, backend/ml/
The decision	For every ping: let it through, hold it until a checkpoint, or put it in a digest. Trusted senders always pass. When evidence is thin it asks you instead of deciding.	/api/simulations, Smart inbox
The feeling	Makes the number felt: swipe cards, a jar of held pings, doors that open onto comic futures, texts from future you, haptics for every choice. Comedy on top, real numbers underneath.	frontend/src/components/*

It runs as a website, an installable PWA, and Android/iOS apps (Capacitor) from one React codebase.

See it on a phone in 15 seconds

Open /demo (for example http://localhost:5180/demo). It's a self-playing phone:

1. You're deep in DBMS notes, focus timer running.
2. An Instagram notification lands and the phone buzzes.
3. Parallel's strip appears on top of the notification: “−18 min if you tap · 74% rabbit hole”, with Peek / Hold / Open.
4. Peek: Door A shows The Scrollfather; Door B shows Main Character.
5. You hold it. It drops into the jar until 1:00 PM; focus keeps going.
6. At 1:00 PM the held pings arrive as one calm batch, and future you texts back.

Simulations: five real-life scenarios

Signed in, open Simulations (/app/simulations); without an account, /demo. Pick a moment and watch it play on a phone. Each one is also a ready-made 1080×1920 video:

Scenario	What it shows	Video
Study session: the Instagram like	Cost on top of the notification, both doors, the jar, the 1 PM batch	study.mp4
Developer in flow: three pings, three calls	Algorithm bait held, Mom (VIP) passes, PR review to the 12:30 digest	dev.mp4
1 AM doomscroll: the late-night guard	Night guard, Door A's worst case, future you, wind down, morning digest	night.mp4
Exam week: the group-chat storm	Strict mode, 14 messages → 1 held thread, auto-reply draft, 7 PM summary	exam.mp4
The “we miss you” trap	Manipulation X-ray, teaser unmasking, re-engagement shield, aura	reengage.mp4

/demo?scenario=<id>&clean=1 is the recording mode (just the phone and captions) used to make these videos; scenarios are data in frontend/src/lib/scenarios.ts, so adding one is a few dozen lines.

Feature tour

The twin learns everything the problem statement asks for

The twin learns…	How	Where you see it
When you focus best	Focus quality by weekday × hour, shifted by your declared wake/bed time and chronotype	Fingerprint heatmap, 3D ring, golden hours
What causes interruptions	Calibrated p(open | context): task depth, fatigue, time, sender, recent switches	Multiverse “why” factors, Twin brain
Which content causes prolonged sessions	p(rabbit hole) plus dwell-time models per app/category	Fragmentation bars, Replay
Which notifications matter	Your “did this matter?” taps, rules and randomized uplift per category	Smart inbox, Today
When you're fatigued	Behavioural estimate plus your 10-second energy check-ins	Today, planner, coach
Which apps fragment you	Breaks per day × cost per break; tool apps vs nag apps	Fingerprint, attention debt
“If I allow this now?” + counterfactual	2,000 Monte Carlo futures; randomized holdouts; T-learner	Multiverse (hero feature)

Personas: the twin's starting guess

On day one the twin has no history, so it borrows the typical pattern of your kind of busy, then moves towards you as your data arrives. Its personal weight grows as n/(n+20).

Persona	Starting focus peaks	Work tools	Most frequent pings	Fatigue usually from
Student	10 AM and 9:30 PM	Notion, Docs, lectures	Social likes, DMs	4 PM
Developer	11 AM and 11 PM	VS Code, Terminal, GitHub	Work chat, email, news	4:30 PM
Creator	2 PM and 8 PM	Figma, Docs, Notion	Social likes (their feed is also their work)	6 PM
Professional	9:30 AM and 3 PM	Excel, Docs, Notion	Work chat, email, meetings	3 PM

Pages

Page	What it does
Today	Echo the 3D companion, focus capacity, fatigue, golden hours, energy battery check-in, “did this matter?”, knowledge meter, held-pings jar, texts from future you, attention budget, focus reservation, context checkpoint, quests
Multiverse	Pick a ping and the moment it arrives. Guess its cost first, then see two comic universes over grounded numbers, raining futures, dominoes, the Door to the Future, a drag-to-fork decision, a holo badge card and a reality check that grades the prediction
Smart inbox	Swipe deck on phones (right allow, left hold, up digest, down noise), list on desktop, Peek on every card, deferred queue with a pourable jar, auto-reply drafts in 4 languages, and a till receipt when you clear it
Fingerprint	Your attention heatmap, 3D ring and fragmentation by app
Replay	Your day as a timeline of switches and focus blocks
Twin brain	Explorable knowledge graph of tasks, times, triggers, apps and outcomes
Accuracy	Calibration, comparison against a generic rule, drift (PSI), prediction-vs-reality
Data & models	Dataset preview, model registry, metrics vs baselines, hash-chained audit log
Coach (copilot)	Answers with cards you can act on: attention receipts, ping prices with Peek / Hold / Open, plans, the jar, X-rays. Acts for you with Undo (“hold everything except Mom for 90 minutes”, /hold 2h, /focus 50), diagnoses “I can't focus”, nudges you at most 3 times a day, remembers commitments. See docs/COACH.md. Voice and text in English, हिन्दी, मराठी, Español. Answers 25+ kinds of question (what now, plan my evening, Instagram vs YouTube, time lost this week, doomscrolling, exams, sleep, stress…), remembers the conversation and offers follow-up chips. Every number comes from a facts sheet built by your models; with an ANTHROPIC_API_KEY on the server, Claude writes the reply from the same facts
Lab	Attention budget and debt, co-focus rooms, habit experiments, developer Attention API + JS SDK, team health (aggregates only, k ≥ 5)

It knows your day

• 7-step onboarding: persona and goals → wake/bed time and chronotype → weekly schedule (templates) → deadlines and exams → tool vs nag apps → who always gets through. Schedule and deadline steps can be skipped; the same editor lives in Settings.
• Schedule-aware decisions: during a class or work block, low-value pings are held and the reason names the block (“You're in Lectures until 1:00 PM…”). A deadline within 3 days makes the twin stricter.
• Best moments to check held pings, from when your blocks end. On native these become local reminders.

Trust and privacy

• Metadata only: message text and window titles are never stored.
• Pause learning, set a retention window, export a portable profile, delete everything.
• Gamification without guilt: XP for intentional choices (allowing on purpose counts), anti-farming rules, a daily cap on holding back, no shame mechanics.

The phone-first, Gen Z layer

The core numbers are serious; the way you meet them doesn't have to be. Everything below is built, tested at 375 px, and documented in docs/PHONE_EXPERIENCE.md.

Feature	What it feels like
Swipe triage deck	Tinder for pings. Right allow, left hold, up digest, down “noise” (also teaches the twin it didn't matter). Stamps, a price tag in focus minutes, and a distinct haptic per direction.
Held-pings jar	A physics jar. Every hold drops a ball in, coloured by category. Tap or shake the phone to pour them out as one calm batch.
Drag to fork	Drag a token into Universe A or B. Live “+min” counters run as you drag; it ticks every quarter of the way.
Energy battery	Drag a battery up or down to check in. One tick per level.
Guess before reveal	Bet how many minutes the ping would eat before the twin answers. “Ate.” / “Delulu alert.” / “Dramatic much?”
Raining futures	The 2,000 simulated futures fall as dots and pile into the cost histogram.
Dominoes	Universe A's chain as dominoes sized by the minutes each step adds. They fall, one haptic each.
Holo badge cards	Each decision earns that universe's archetype card (Main Character Energy, Professional Thumb Athlete…). The foil follows your finger or the gyroscope. 12 to collect in a binder.
Inbox receipt	Clear the inbox and a till receipt prints: focus protected, opened on purpose, XP, “NO REELS WERE HARMED”, Share.
Door to the Future	Two doors per ping. Each opens onto a four-panel comic drawn in code from your numbers, ending on The Scrollfather or Main Character, with a shareable 1080×1920 poster. See docs/FUTURE_PEEK.md.
Manipulation X-ray	Each ping's wording is scanned (in memory only) for social proof, vague teasers, fake urgency, re-engagement bait, streak pressure, algorithm bait and FOMO hooks. Tap a chip to see how the tactic works; vague pings get “unmasked” with what they usually turn out to be.
Shields	The re-engagement shield sends “we miss you” pings to the digest; the holds everything but VIPs from 30 minutes before bed until you wake.

A single haptic language runs through all of it (frontend/src/lib/feel.ts): native Capacitor haptics in the app, navigator.vibrate on Android Chrome, silent elsewhere, and a parallel.haptics = off switch.

How it would run over real notifications

Today the inbox is fed by realistic sample pings. The production design is documented step by step in docs/REAL_NOTIFICATIONS.md. In short:

• Android (the real target): a NotificationListenerService (the user grants Notification access) sees each notification's package, category and timing, drops the text, asks the twin, and either lets it through, snoozes/cancels it and re-posts it at the checkpoint, or adds it to the digest. A heads-up “Parallel” notification with Peek / Hold / Open actions is how it appears on top of the original. Usage access (UsageStatsManager) gives the app timeline.
• iOS: apps cannot read other apps' notifications. Parallel works through Focus filters, Screen Time (FamilyControls) and its own inbox.
• Desktop: ActivityWatch today (POST /api/ingest/activitywatch); Windows UserNotificationListener later.

The API reports this capability matrix at GET /api/platform/capabilities.

Data & ML (honest version)

• python -m ml.dataset creates a synthetic population: 60 users × 30 days, 81,582 notification events × 35 columns and 149,649 app sessions, saved as CSVs and SQLite tables.
• python -m ml.train trains population models evaluated on users never seen in training:

Target	Model	Unseen-user result	Baseline
p(open)	Gradient boosting + isotonic	AUC 0.845	0.728
p(rabbit hole)	Gradient boosting + isotonic	AUC 0.802	—
Focus cost p10–p90	Quantile GBMs	79% coverage, MAE 5.5 min	MAE 6.0
Counterfactual	T-learner	0.9 min gap vs randomized truth	—

• Every account's history lives in SQLite (user_segments, user_notifications). Real events (source='live') are appended, and the twin refits from the database.
• Details: docs/DATA_AND_ML.md.

Quick start

git clone https://github.com/vizsh/Parallel.git
cd Parallel
./run.sh

On Windows: powershell -ExecutionPolicy Bypass -File .\run.ps1. Then open http://localhost:5180 (API on 8787). The script creates the Python venv, installs both halves, builds the synthetic dataset and trains the models on first run.

Demo login	Password
[email protected] · [email protected] · [email protected] · [email protected]	parallel123

On your phone (same Wi-Fi, no installs): start Vite with --host (npm run dev -- --host), open http://<laptop-LAN-IP>:5180 and use Add to Home screen. The dev server forwards /api to the API on your laptop, so don't set VITE_API_URL for this: a hard-coded IP or port breaks every API call as soon as either changes. Allow Python and Node through the firewall on private networks if the phone can't connect. Full instructions: RUN.md and docs/MOBILE.md.

APK: needs the Android SDK (it comes with Android Studio): cd frontend && npm run android, then Build → Build APK(s). Java alone isn't enough, because Gradle needs Android's platform jars and build tools to compile and package the native shell.

Architecture

React 19 (web · PWA · Capacitor Android/iOS) ──HTTPS/JWT──▶ FastAPI ──▶ SQLite
   3D landing (R3F + GLSL) · 12 app screens        twin.py · routine.py · population.py    users · profiles · history
   phone layer: swipe · jar · doors · texts         routine_api · platform_api · assistant   predictions · audit chain
   i18n (4 languages) · voice coach · haptics       ▲ versioned models (ml/train.py)         ds_* population dataset

The life of one ping, end to end, is in docs/ARCHITECTURE.md and docs/CONTEXT.md.

Layer	Stack
Frontend	React 19, Vite 6, TypeScript, Tailwind v4, motion (Framer Motion), three.js + React Three Fiber, GSAP, react-i18next
Native	Capacitor 8 (Android, iOS): haptics, local notifications, status bar, app
Backend	FastAPI, SQLite, scikit-learn, NumPy/pandas, JWT auth
ML	Gradient boosting + isotonic calibration, quantile GBMs, T-learner, Monte Carlo rollouts

Repository map

Parallel/
├─ README.md · RUN.md · CHANGELOG.md · CONTRIBUTING.md · run.sh · run.ps1
├─ docs/
│  ├─ CONTEXT.md             ← start here: the whole project's context, decisions and gotchas
│  ├─ LOGIC_AND_FORMULAS.md  every formula, transcribed from the code
│  ├─ PSYCHOLOGY.md          notification manipulation, the research, and Parallel's counter-design
│  ├─ PHONE_EXPERIENCE.md    every playful phone interaction, how it works, where it lives
│  ├─ FUTURE_PEEK.md         Door to the Future + Texts from future you + the Gen Z backlog
│  ├─ REAL_NOTIFICATIONS.md  running over real notifications: Android, iOS, desktop, permissions
│  ├─ ARCHITECTURE.md · DATA_AND_ML.md · MOBILE.md · API.md · FEATURES.md · ROADMAP.md
│  ├─ IDEATION.md · PRESENTER_GUIDE.md · Parallel_Project_Report.(pdf|docx)
│  └─ screenshots/
├─ sdk/parallel-sdk.js                  Attention API client
├─ backend/
│  ├─ app/   main · twin · routine · routine_api · platform_api · population · history · synth · assistant · gamify
│  ├─ ml/    dataset.py · train.py
│  └─ tests/ test_api.py
└─ frontend/
   ├─ src/landing/     TwinWorld (WebGL world) · shapes · scroll
   ├─ src/pages/       Landing · Auth · Onboarding (7 steps) · Demo · Privacy · Contact · NotFound
   ├─ src/app/         Today · Multiverse · Inbox · Fingerprint · Replay · Brain · Lab · Accuracy · Data & models · Coach · Settings
   ├─ src/components/  SwipeDeck · HeldJar · ForkDrag · EnergyBattery · GuessGame · FuturesRain · Dominoes · BadgeCard
   │                   Receipt · FuturePeek · ComicArt · FutureTexts · PhoneDemo · Universes · RoutineEditor · VoiceCoach · charts
   ├─ src/lib/         api · auth · feel (haptics) · future (comic builder) · futureTexts · cards · universes · native · i18n
   ├─ android/ · ios/  Capacitor native projects
   └─ public/          PWA manifest, service worker, icons, robots, sitemap, OG image

What is built, what is simulated, what is next

	Status
Personal twin, Multiverse, Smart inbox, population models, registry, audit	✅ built and working
Routine-aware decisions, onboarding, check-ins, check moments	✅ built
Phone-first layer (swipe, jar, fork drag, battery, guess, rain, dominoes, cards, receipt)	✅ built, tested at phone width in a browser
Door to the Future, Texts from future you, /demo	✅ built, tested in a browser
X-ray, shields, aura & fate, peephole, plot twist, Simulations (5 scenarios + videos), rebuilt coach	✅ built, tested in a browser
Haptics, shake and gyro on a real device	🟡 code in place; vibration works in Android Chrome; shake and tilt need HTTPS or the native app; not yet tested on hardware
Inbox pings	🟡 realistic samples from the API, not your phone's real notifications
Day-0 history and accuracy numbers	🟡 synthetic, shaped by your declared routine; real data replaces it as it arrives
Badge cards and future texts	🟡 stored on the device (localStorage), not synced to the backend
Android NotificationListener plugin	🗺️ next; fully specified in REAL_NOTIFICATIONS.md
Signed APK / store builds	🗺️ needs Android Studio (SDK) or a CI build

The ordered plan is in docs/ROADMAP.md; the Gen Z feature backlog (aura points, plot-twist panels, weekly webtoon, door skins, co-op comics…) is in docs/FUTURE_PEEK.md.

Documentation index

Doc	Read it when you want…
docs/CONTEXT.md	the full story: what exists, why it was built this way, conventions, gotchas, how to continue (also written for AI coding agents)
docs/FEATURE_LIST.md	every feature in one place, how it works, and its status
docs/LOGIC_AND_FORMULAS.md	every formula behind the numbers: the Monte Carlo chain, shrinkage, the decision rule, confidence, blending, evaluation, XP
docs/PSYCHOLOGY.md	how apps use notifications to pull people in, the evidence, and how each Parallel feature counters it
docs/PHONE_EXPERIENCE.md	how each interactive phone component works and where its code is
docs/FUTURE_PEEK.md	the Door to the Future spec, the comic system, future texts, the Gen Z backlog and art references
docs/REAL_NOTIFICATIONS.md	how Parallel would sit on top of real notifications on Android, iOS and desktop
docs/ARCHITECTURE.md	the system diagram and request flow
docs/DATA_AND_ML.md	the dataset, features, models and evaluation
docs/MOBILE.md

Honest limits

• Day-0 history is generated. It follows your declared routine, but it is not behaviour you actually showed. Settings says so, and real data replaces it.
• Accuracy numbers are on synthetic data. The best upgrade is a few days of your own ActivityWatch data (POST /api/ingest/activitywatch?tz_offset_min=330; ActivityWatch times are UTC, 330 is IST).
• iOS can't read other apps' notifications. Android and desktop are the realistic first platforms.
• Fatigue is a behavioural estimate, never a medical claim.
• The comics are comedy over estimates. Every one says so: “an estimate, not a prophecy”.