EmoPulse - Under the Hood

Section I

The Five-Layer Architecture

Every AI response is a function of the current multimodal emotional state, filtered through mathematically precise risk assessment and behavioral policy. The loop never stops.

Layer I - Signal Acquisition

Multimodal Signal Acquisition

Three independent channels capture raw physiological truth simultaneously. Camera extracts heart rate via rPPG, micro-expressions, gaze, breathing. Text engine reads tempo, pressure, escalation. Voice analyzes pitch, pauses, tension. Each produces an independent feature vector.

V_A (biometric) · V_B (linguistic) · V_C (prosodic)

rPPG 478 Landmarks FACS Action Units Gaze Tracking Voice F0 Semantic Pressure

✓ BUILT - Camera + Text engines operational

Layer II - Feature Computation

47 Normalized Parameters

The camera stream (V_A) is computed into 47 normalized parameters across emotion, physiology, and behavior. Voice and text (V_B, V_C) stay as separate vectors, fused in at Layer III. Proprietary ontological processing. Each parameter is intensity-scored, stability-tracked, and trend-directed.

O(t) = Φ(Es(t), K) - where K is the knowledge base

RDF/OWL 500+ Nodes Plutchik Wheel Parrott Tree OCC Model

✓ BUILT - Full ontology operational

Layer III - Cross-Stream Coherence

Dynamic Fusion & Coherence Validation

The core of EmoPulse. Signals from multiple streams are validated for coherence via dynamic weighted fusion - weights computed at runtime based on signal availability and quality. Cross-stream contradictions are detected and resolved before state assembly.

v = α·V_A + β·V_B + γ·V_C → R_risk = f(v) → π(v)

R_risk < 0.30 → NORMAL | 0.30-0.60 → ELEVATED | 0.60-0.85 → DIRECT | ≥ 0.85 → MINIMAL

Dynamic α,β,γ 6-Dim State Vector Risk Scalar 4 Control Modes Patent Pending

✓ BUILT - Full decision pipeline operational

Layer IV - State Vector Assembly

R_risk, intent_clarity, human_signature

Coherent signals assemble into a state vector: R_risk (risk scalar), intent_clarity, and human_signature. The risk scalar partitions the continuum into four operational modes. Each mode generates a structured control policy that constrains AI output via pre-generation, live enforcement, and post-generation correction.

response = AI(input, π) - where π constrains generation

Pre-generation Live Enforcement Post-generation Check

✓ BUILT

Layer V - Temporal Context

Personal Baseline + Deltas

Every output becomes the next input. Personal baselines track individual norms; deltas detect deviation from the person's own patterns. Coherence monitoring detects semantic degradation via divergence measurement. Long-term memory persists across sessions on the device. The loop never stops - no discrete end state.

D(t) = |O_pred(t) - O_obs(t)| > θ → F(t+1) = αF(t) + (1-α)O(t) + βH(t)

M(t) = γM(t-1) + (1-γ)O_summ(t) - long-term memory

<50ms Cycle Zero Network Coherence Monitor

✓ BUILT

Fusion happens at the interpretive control layer - outside and above the AI model.
Not inside the neural network. Not at the embedding layer. Not predetermined at training time.
Dynamic. Interpretable. Auditable. Runtime-adaptive.

<50ms

On-Device Latency

Parameters Per Frame

Patents Pending

Cloud Dependency

Section II

The Machine Cannot See The Room

Every system that works with a person today is working from a description of that person, not from the person. This is the gap the architecture exists to close, and it is easiest to explain in the founder's own words.

The argument

FOUNDER

"Right now a system works like a robot vacuum cleaner. It has a map, a set of instructions that say if the person is sad be gentle, if angry do not argue. It drives around that map hoping it does not hit the furniture. But it cannot see the room."

FOUNDER

"The point is the architecture, not the demo. The demo only shows that the process works. Think about what changes if the machine has the parameters: it does not have to guess how to act. Not from a list somebody wrote for it, but from what it actually sees in the person."

What that costs in practice

A pilot writes that everything is fine while the pulse says otherwise. An operator reports ready at the end of a fourteen-hour shift because that is what one says. A patient answers a questionnaire the way they think they should answer it. None of these people are lying. They are describing themselves, and a description is a lossy copy of a state.

Text, forms and self-report are the only inputs most systems have. The signal that would have told them what is actually happening is sitting in the camera that is already pointed at the person, unread. Reading it, on the device, without sending an image anywhere, is the whole of what we do.

A machine with access to everything ever written still guesses at the one thing in front of it. It has instructions. It has search. It has analysis. It does not have eyes.

Section III

The Builder

Arvydas Pakalniskis

Founder & CEO · Sole Inventor · ProBuggy #5

Before EmoPulse, my life was measured in dust clouds, broken bolts, and the kind of adrenaline that doesn't politely knock - it kicks the door in and asks if you're awake. I raced in U.S. off-road series like LOORRS and AMSOIL Championship, driving the ProBuggy #5. People think racing is about speed. It's not. Speed is the easy part. The real game is discipline - the kind you only learn when a tiny mistake costs you a wheel, a race, or a few ribs.

Racing teaches you strange things. You learn to read terrain the way some people read books. You learn to feel vibration patterns through the steering wheel and know instantly if something is wrong. You learn that chaos has rules - and if you don't respect them, the track educates you fast. And yes, the adrenaline is real. It doesn't "flow." It floods. But underneath that noise, there's structure. Always.

Years later, when I started building EmoPulse, I realized something funny: the same instincts that kept me alive on the track were the ones helping me design interpretive architectures. Pattern recognition. Signal extraction. Human-in-the-loop discipline. Understanding that systems fail not because they're slow, but because they drift, lose meaning, or collapse under noise. Racing taught me to see noise differently - not as a problem, but as information. And once you learn to read noise, you can build systems that survive it.

So no - EmoPulse didn't come from a lab or a whiteboard. It came from real dirt, real risk, real consequences, and a very real ProBuggy #5 that didn't care about your excuses. That world shaped how I think, how I build, and how I navigate high-stakes AI today. Because whether it's a racetrack or an AI system, the rule is the same: if you don't respect structure, the environment will teach you the hard way.

RACING LOORRS · AMSOIL Championship · ProBuggy #5

PATENTS 3 patents pending

STACK Ontological AI · rPPG · RDF/OWL · On-device ML · Edge Inference

SYSTEM 5-layer pipeline · <50ms latency · 100% on-device · Privacy by architecture

STATUS Live prototype · Internal measurements only, not independently evaluated · Optional server-mode for enterprise audit · Integration work under way with partners in five countries

Advisory Board

Dr. Anastasia Vasina

MD, PhD (Pathology) | Fractional Chief Medical Officer | ex-CPO/CMO, Soter Analytics

Angelo Arcadu

Public Affairs & Strategic Communications | Project Delivery in Defense, Healthcare & Critical Infrastructure | Italian MoD (NATO-aligned) NATO NCAGE & UN UNGM Registered Professional

Why it runs on the device

The regulated domains are closing the door on sending biometric signal to a server. The EU AI Act restricts emotion inference in workplaces and schools. Health and defence procurement rules treat transmitted biometric data as a liability before they treat it as a feature.

Our answer is architectural rather than contractual. In the strict configuration the computation stays on the device: there is no server copy to attack, to subpoena or to leak, and the system keeps working where the link is degraded or denied, because a link was never needed. Where a second person has to see the state in real time, a clinician during a consultation for example, the state travels to that instance under consent, with audit and isolation. The regime is a property of the deployment, not a promise about it, and you can check it yourself: open the network panel on our signature build and count the requests.

You've seen what it does.
Now see how it thinks.

The Machine Cannot See The Room

Why it runs on the device

Every AI will need to see the human it serves.

You've seen what it does.Now see how it thinks.

The Machine Cannot See The Room

Why it runs on the device

Every AI will need to see the human it serves.

You've seen what it does.
Now see how it thinks.