A.R.I.A. — Affective Reasoning & Intelligent Adaptation
Most Wearables
Measure Stress.
None of Them Know You.
538 million devices ship the same one-size-fits-all model. We built the personalization layer that makes it accurate — for each individual. A.R.I.A. is the missing API between raw wearable data and real human-state intelligence.
89.8%
Lab balanced accuracy after per-user calibration (WESAD, N=15, binary)
+10.4pp
Lift over the 79.4% zero-shot baseline (lab)
~56%
Field balanced accuracy at 30 self-report labels (DAPPER, N=20 at that budget, N=84 cohort)
437
Automated tests, ML leakage audited
Galaxy Watch
Validated on Galaxy Watch 5 data (GalaxyPPG, N=23, 71.8%). Live raw-PPG support today: Galaxy Watch 8.
Lab-validated under controlled conditions. Paper submitted to ACM IMWUT, May 2026. Full methodology and data
The Discovery:
Labels, Not Models.
Every competitor is building better algorithms. Our data shows that's the wrong problem. In field conditions, the architectures we tested converge. The bottleneck is calibration — how you collect and label the data.
Finding 1
Architecture Converges
In field conditions, the three architectures we tested converge between 0.534 and 0.565 at 30 labeled moments (DAPPER, N=20 at that budget). No pairwise comparison reaches significance after correction, and no large effect size appears across 21 tests. Medium effects cannot be ruled out at this sample size.
Finding 2
Calibration Changes Everything
A short structured calibration session covering both activated and resting states lifts balanced accuracy from 79.4% to 89.8% in controlled conditions. That's the gap between a generic stress score and one that actually knows you. Personalization is the product.
Finding 3
First to Measure It
No prior paper compares calibration cost across model families under identical protocols. Our 25-page paper (submitted to ACM IMWUT, May 2026) is the first systematic measurement. This reframes the field.
Population-level models (Apple, Garmin)
“Stress: 87%”
Black-box algorithm
Same model for everyone
No personalized calibration
No per-user explainability
A.R.I.A. — Calibrated to you
“Elevated arousal — calibrated to your baseline. Consistent pattern during evaluative contexts.”
Explainable — shows why, not just what
Personalized to each individual
Private by architecture — models small enough to run on-device
Transparent and auditable
Three Layers,
Each Independently Valuable.
A.R.I.A. is not a single product. It is a progressively richer system where each layer's data builds the next.
Layer 1 — Now
Wrist: Arousal Detection
Stressed or not stressed — detected from your existing wearable, calibrated to your own baseline in a short structured session covering both activated and resting states. Already works in controlled conditions.
Every competitor ships a generic model. This is the first personalized one.
Layer 2 — Next
Wrist + Voice: Emotion
Add voice analysis (opt-in) to distinguish happy from stressed, sad from calm. Custom labels that learn your emotional vocabulary over time.
No competitor combines wrist calibration with voice.
Layer 3 — Destination
Your Emotional Map
Over time, the system builds a map of your emotional life that is unique to you — discovered from your own data, not predefined categories.
Nothing like this exists from wearables today.
The Destination
Beyond Emotion — The Platform
A.R.I.A.'s calibration methodology is state-agnostic. The same pipeline that personalizes arousal detection extends to focus, fatigue, cognitive load, and any state a user can label. We start with affect because it's the strongest wrist signal. The platform grows into a unified human-state API — where arousal, emotion, attention, and energy are different views of the same per-user model.
Focus & Attention
Via future EEG integration
Fatigue & Energy
Via longitudinal wrist patterns
Cognitive Load
Via multimodal sensor fusion
Custom States
“Deep work,” “pre-meeting dread,” “recovery” — user-defined, system-learned
Same calibration methodology. Same privacy architecture. Same API. Every new state makes the per-user model richer.
538M
wearables shipped in 2024 — sensor infrastructure already in pockets
$17B+
TAM by 2030 — digital therapeutics and health APIs ($8B, Grand View 2024), XR biometric tooling ($6B, ARinsider/IDC 2026), enterprise productivity and wellness ($3B, McKinsey 2024)
$3B
SAM by 2027 — the serviceable slice of those segments where a per-user calibration API fits today
Built by Scientists
and Engineers.
437 tests. 25-page paper submitted May 2026. Real-time streaming API. Galaxy Watch validated. This team ships.
Research Advisor
Sonia
Litwin
10+ years EEG/HRV research.
CTO & Co-founder
Marco
Accardi
Built the research pipeline and calibration architecture. Sole inventor on the USPTO provisional.
CEO & Co-founder
Alessandro
De Angelis
Leads strategy, partnerships and US go-to-market.
CMO
Sabrina
Pippa
Brand and visual communication.
Partners
Oracle for Startups · NVIDIA Inception
For Researchers
Run A.R.I.A.'s calibration analysis on your own wearable dataset — at no cost. Get per-subject accuracy with and without personalization, calibration gain analysis, and architecture comparison.
Learn MoreFor Device Makers
Building a wearable with stress detection? Our submitted paper measures what per-user calibration costs on Samsung Galaxy Watch data — the first systematic measurement across model families under identical protocols.
Get in Touch