Meet SANSARA

Name: Sansara
Price: 29.00 USD
Availability: PreOrder
Author: Sansara

a quiet kind of empathy

Nine AI agents. On-device inference. Zero cloud. A 3×3 empathy matrix that debates, critiques, and refines every response before it reaches you — entirely inside your phone.

Coming March 2026

Read the deep-dive↓Get Notified

The Problem

Every wellness app lies to you.

Your data on their servers

You write your most vulnerable thought at 2 AM. It gets encrypted in transit, sure — then stored in plaintext on a server farm in Virginia. Searchable. Subpoena-ready. Breach-waiting.

One model, one perspective

Every other wellness app runs a single language model with a single system prompt. One perspective. No cross-checking. No debate. It hallucinates a platitude and calls it therapy.

Monthly rent on your mental health

$14.99/month. $179/year. For access to your own emotional data. Miss a payment and your journal history goes dark. Your wellness held hostage by a billing cycle.

Analytics on your vulnerability

They track which entries make you cry. How long you linger on anxiety screens. Your emotional metadata — packaged, anonymized, sold. Your pain is their product metric.

We didn't build a better version of this. We built something structurally different.

Philosophy

Not a chatbot. A presence.

Unbothered Presence

Sansara doesn’t panic when you do. It holds space without rushing to fix, because sometimes being heard is the whole point.

Reflective, Not Prescriptive

Instead of telling you what to do, Sansara helps you see what you already know. Gentle mirrors, not instruction manuals.

Anti-Toxic Positivity

No "just think positive!" platitudes. Sansara sits with discomfort, validates darkness, and trusts your own timeline for healing.

Growth Without Gamification

No streaks. No guilt. No points. Just constellations that quietly map your journey — visible only when you want to look.

Experience

What it feels like

I don’t know why I feel so off today. Nothing bad happened, I just... can’t settle.

Sansara

That restless feeling that doesn’t attach to anything specific — it’s real even without a reason. You don’t need to justify it. What does ‘settling’ usually feel like for you?

I guess... quiet? Like my brain stops scanning for threats.

Sansara

That scanning quality — your nervous system doing overtime even when the coast is clear. That’s exhausting, and noticing it is genuinely something. Would it help to explore what usually shifts that for you, or would you rather just sit with it for now?

Every other app vs. Sansara

Generic Bot

“Have you tried thinking more positively? 1. Practice gratitude 2. Go for a walk 3. Try deep breathing”

Sansara (9 agents)

“You don't need to unstick yourself right now. Let's just sit here.”

Real response style · Generated 100% on-device

The Nine

Named. Specialized. Debating inside your phone.

The Sansara Empathy Matrix is a 3×3 grid of AI agents organized across three layers of psychological support. They share a single quantized model but operate under distinct personas, system prompts, and behavioral constraints. Only the agents relevant to your current state activate — selected by a sub-100M parameter router in milliseconds.

Layer 1 — Feeling what you feel

EMOTIONAL MIRRORING

ECHO_EMPATH

Validates your immediate emotional state

Analyzes real-time text sentiment and vocal prosody to mirror what you’re feeling — without rushing to fix it. If you’re frustrated, it says so. If you’re grieving, it sits with you.

PATTERN_ECHO

Recognizes recurring emotional themes

Cross-references your current input against months of local memory. Surfaces patterns you can’t see from inside the moment. “This reminds me of the stress spike you logged last month.”

BIO_MIRROR

Connects your body to your mood

Integrates passive sensor data — step count, ambient light, heart rate variability — to reflect the link between physical state and emotional weather.

Layer 2 — Helping you think deeper

COGNITIVE REFRAMING

INSIGHT_WEAVER

Transforms negatives into reflective prompts

Takes "I failed the presentation" and restructures it into "What hidden strength emerged during that difficult moment?" Cognitive shifts without condescension.

TREND_SAGE

Finds positive arcs in months of data

Analyzes historical vector embeddings to reveal macro-trends hidden within micro-failures. “Over the past six months, these dips always precede your biggest breakthroughs.”

ENV_THINKER

Links environment to cognitive state

Reframes your mental state using physical context — ambient light, location patterns, time of day. “Your sensors show dim light for 48 hours. A brief walk outside might shift this lens entirely.”

Layer 3 — Moving you forward

COMPASSIONATE ACTION

ACTION_NURTURER

Immediate therapeutic micro-interventions

Detects acute stress peaks and offers targeted grounding. “Your vocal stress is peaking — breathe with me for exactly 30 seconds.” Calibrated to urgency, not generic advice.

GROWTH_GUIDE

Personalized strategies from your history

Draws on your database of successful past interventions. Recommends what has demonstrably worked for you — not what worked for a statistical average.

HABIT_ELEVATOR

Behavioral change tied to physical metrics

Links therapeutic action to biometric telemetry. "You haven’t hit your sunlight streak in three days — can we link a five-minute walk to a quick mood lift?"

The Blackboard

Propose. Critique. Refine. Converge.

Agents don't talk to each other. They read and write to a shared Blackboard — a mutable JSON state object holding your raw input, historical embeddings, active hypotheses, and intermediate proposals. This decoupled architecture eliminates the “Phone Game” problem where context degrades through sequential handoffs. Every agent sees the full picture. Every critique is grounded in the same data.

Propose

Each activated agent analyzes the Blackboard state and writes an independent response proposal. No agent knows what the others are writing.

Critique

Agents read each other’s proposals from the shared Blackboard. They post structured critiques — identifying blind spots, cognitive distortions, or mismatched intensity.

Refine

Each agent integrates peer feedback and rewrites its proposal. The loop repeats until semantic convergence is measured across all active outputs.

Converge

The Control Shell detects mathematical consensus — output vectors stabilize within threshold. One unified response emerges. Sub-second, on-device, invisible to you.

the vortex

Agents Synthesizing

validating…reframing…historical match foundsynthesizing…consensus reached

Engineering

The stack that makes this possible.

50KBruntime footprint

ExecuTorch

Meta’s ahead-of-time compilation framework. Captures the PyTorch computational graph via torch.export(), quantizes it, and lowers it to hardware-specific backends. The output is a .pte binary that runs directly on silicon. No interpreter. No VM. Math on metal.

~1.5GBtotal RAM for inference

4-Bit Quantization

Sub-billion parameter models compressed from 16-bit to 4-bit precision using SmoothQuant and SpinQuant rotation matrices. 70%+ memory reduction while preserving reasoning capability. Fits comfortably on 6GB+ devices.

3MBdatabase binary

ObjectBox HNSW

On-device NoSQL vector database with Hierarchical Navigable Small World indexing. Millions of semantic embeddings searched in milliseconds via memory-mapped C++ core. Zero cloud. Zero latency.

2.2–3.6×inference speedup

Speculative Decoding

A lightweight draft model predicts multiple future tokens simultaneously. The primary model verifies them in one batch operation. Bursty inference: max power for a fraction of a second, then back to sleep. Battery-friendly.

// Hardware acceleration targets

iOS → CoreML backend → Apple Neural Engine

Android → QNN backend → Qualcomm AI Engine / Adreno GPU

Fallback → XNNPACK → Optimized CPU inference

// Model loading

const model = ExecuTorch.load("gemma-3-1b-q4.pte");

const memory = ObjectBox.encrypted("AES-256-on-device");

// Network access: physically impossible.

Multimodal

It hears the stress in your voice.
It sees the tension in your face.

Text alone can't capture how you feel. Sansara fuses three sensory modalities into a unified context vector before routing it to the Blackboard. Every sensor is opt-in, per-session, and the raw data is purged immediately after embedding.

Voice Prosody Analysis

Whisper ASR · Tiny.en · ~40MB RAM

Sansara doesn't just transcribe your words — it listens to how you say them. Speech pacing, pause frequency, and token density are analyzed to build a prosody vector that captures emotional states beyond language.

0:00stress detected: elevated pacing0:12

0.72

tension

0.18

calm

0.61

focus

Facial Expression Recognition

ML Kit + MobileNetV3 · ~5MB RAM · Real-time

A two-stage pipeline detects facial landmarks, then a quantized classifier outputs a 192-dimensional emotion probability vector. Raw pixel data is purged from RAM instantly — only the embedding survives, and it never leaves your device.

Biometric Telemetry

HealthKit · SensorManager · Passive

Background polling of step count, ambient light, and heart rate variability. Establishes a physiological baseline that contextualizes your emotional state with physical data. Zero active interaction required.

steps: 2,847

light: 12 lux (dim)

hrv: 34ms (elevated stress)

sleep: 5.2h (below baseline)

→ BIO_MIRROR + ENV_THINKER activated

Continuous Learning

The model becomes yours.

While you sleep, Sansara runs Low-Rank Adaptation on your device. LoRA freezes the base model's weights and injects tiny trainable matrices — targeting the attention layers where personality lives. Millions of parameters, not billions. Feasible on mobile silicon.

By morning, the model knows you better. The adapter is 10-50MB. Your base model is untouched. No data left your phone. No server was involved.

On Apple Silicon, this runs via MLX Swift. On Android, ExecuTorch handles the training loop natively. A time-series meta-agent also runs ARIMA forecasting on your emotional vectors — predicting mood dips before they hit.

Dedicated Feedback Sessions

Once a week or once a month — your choice — Sansara invites you to a dedicated feedback conversation. Not a thumbs up. Not a star rating. A real dialogue about what landed, what missed, and what you actually needed.

This feedback feeds directly into the on-device LoRA training loop. The result is a model that doesn't just learn from your journal entries — it learns from your explicit corrections. No other wellness AI has this.

“Last Tuesday I suggested a breathing exercise when you were venting. How did that feel?”

“Honestly? I just needed you to listen. The exercise felt dismissive.”

“Got it. When you're in that state, I'll hold space first. Training updated.”

Completely optional. Set your cadence in settings. Between sessions, the model also learns passively from the quality of your responses — longer entries, re-engagements, and natural conversation flow all signal what works.

New

Write it. Draw it. Sansara captures it.

Tablet Mode turns your device into a freeform canvas. Write with a stylus, sketch your thoughts, diagram your feelings — Sansara's on-device recognition converts handwriting and drawings into structured journal entries while preserving the originals.

Tablet Mode

Handwriting Recognition

On-device OCR converts handwritten entries into searchable, embeddable text. The original strokes are preserved alongside.

Freeform Drawing

Sketch emotions, diagram thought patterns, or just doodle. Sansara analyzes visual patterns for emotional context.

Stylus-Native

Pressure sensitivity, palm rejection, and tilt detection. Optimized for Apple Pencil and Samsung S Pen.

Accessibility

Built for every mind.

15-20% of the population is neurodivergent. Most apps treat accessibility as a checkbox. Sansara treats it as a design constraint — every feature is built with ADHD, ASD, and dyslexia in mind from day one.

Eight interface modes from Minimal to Maximum visual intensity

OpenDyslexic font with adjustable letter spacing and line height

Granular haptic, motion, and soundscape toggles for sensory sensitivity

Voice input for hands-free journaling with prosody analysis

ADHD Focus Mode for structured attention without overwhelm

Stellar Alignments — cumulative progress that forgives absence

No streaks, no guilt counters, no shame mechanics

WCAG-compliant glassmorphism with enforced text legibility

Models

Same architecture. Different engine.

Nothing is withheld from base users. Every tier gets the full 9-agent Empathy Matrix, full 5-year memory, full LoRA personalization, full feedback sessions, and full privacy. The only difference is the language model underneath — and that's determined by your device's hardware. If your device can run the bigger model, you get the bigger model. The small price difference covers our cost of storing and distributing larger model files. That's it.

SANSARA

The complete product

Llama 3.2 1B QLoRA

Full 9-agent Empathy Matrix
5-year companion memory
LoRA personalization overnight
Weekly reflection sessions
Voice journaling + prosody
Facial expression + biometrics
Predictive mood foresight
8 interface modes + ADHD Focus
Stellar Alignments
Zero cloud. Zero telemetry.

Everything. No asterisks.

SANSARA Pro

More powerful engine

Llama 3.2 3B

Everything in SANSARA, plus:

Handwriting input — stylus writes, SANSARA transcribes
Freeform drawing — sketch moods and thought diagrams
Apple Pencil / S Pen native

iPad, Galaxy Tab, Galaxy S25 Ultra
Devices with ≥ 6 GB RAM

SANSARA Desktop

The deepest model available

Llama 3.2 11B

Everything in Pro, plus:

Deeper vocabulary, longer context, more nuanced responses
Vision — SANSARA sees your drawings and incorporates them
Split-pane layout — history left, active session right

M-series Mac, iPad Pro M4
Devices with ≥ 10 GB RAM

Your tier is determined by your device, not your wallet. If your hardware can run the bigger model, you get the bigger model. The price difference covers model distribution costs — nothing more.

The Manifesto

No server. No compromise.

We don't have a server to hack. Not because we secured it well. Because it doesn't exist.

The app is physically incapable of transmitting your data. Not by policy. Not by promise. By architecture.

network_permission: NONE

api_keys_in_binary: NONE

analytics_sdk: NONE

cloud_endpoints: NONE

telemetry_hooks: NONE

data_location: YOUR DEVICE

encryption: AES-256

bytes_transmitted: 0

Auditable · Verifiable · Structurally incapable of betrayal