Dramatization · fictional mission

The crew is asleep.

All systems nominal.

Master alarm
Cabin pressure drop · Node 2
03:12:08  Pressure drop detected · node 2
03:12:09  Sealing hatch 2-A
03:12:11  Source: valve V-14 left open
03:12:12  Last operated 01:12 · maintenance shift
Cabin pressure 101.3kPa
Signal lost
Flight recorder · rewinding
−2 h19 hours awake.
−24 h4 hours of sleep.
−48 hWarning signs. Unnoticed.

The signs were there.

Nobody was watching.

PULSAR, a round white robot with a glowing smiling face, waving, circled by an orbit ring
Meet

P. U. L. S. A. R.

Your personal health companion in space.

NASA Space Apps 2026

19 s · sound on · Space pauses

Skip intro →
0:00 / 0:19

Presenter notes

Technical backup / after the pitch

Presenter setup / frontend address

Start the frontend with “pnpm dev” in frontend/ and the API with “uv run uvicorn main:app” in backend/. Ollama and the downloaded models are needed for local AI and voice.

Only pulsar.alexvinola.com, localhost and loopback addresses are accepted. If embedding is blocked by the browser, use the new-tab link and return to this presentation afterwards.

Current HUD: camera indicators and rules are wired in. Sleep, mood and cabin values use nominal context; a live PVT-B screen is not included in this frontend yet.

Open the technical appendixModels · methodology · sources / for questions after the pitch

Models and implementation status

IN CODE

MediaPipe Face + Pose

FaceLandmarker and PoseLandmarker Lite run in a browser Web Worker, with GPU-to-CPU fallback. Eye closures, yawns and posture feed a fixed-weight estimate.

LOCAL OR CLOUD INFERENCE

Gemma 4 E4B · QAT

Default Ollama tag: gemma4:e4b-it-qat. Phrases the rules verdict and supports separate visual descriptions. Sampled frames stay on the laptop; visual agreement is exploratory, not ground truth.

VOICE IN / OUT

Whisper + Piper

faster-whisper base.en uses CPU int8 for transcription. Piper en_US-lessac-medium.onnx synthesizes speech. These defaults are English; bilingual presentation text does not change the voice models.

PLANNED / G2

Trained fatigue classifier

G2 specifies feature extraction, per-person baselines and leave-subject-out validation. No trained fatigue_model.json or G2 metrics artifact is present in this checkout. The current score is heuristic.

How we build and evaluate

Signals are estimates.

The current eye-closure window uses MediaPipe blendshape thresholds. It is an approximation to the PERCLOS concept, not a calibrated measurement of 80% eyelid closure.

Rules are traceable.

Every threshold cites a registered source or declares a team assumption. JSON Schema validates the shared telemetry contract. Scenario tests exercise the decision policy.

Language is bounded.

The voice pipeline checks generated phrasing for selected violations and falls back to a deterministic message when those checks fail. The rules remain responsible for the health verdict.

Validation is separated from the demo.

Subject-separated training is planned for G2. A recorded voice bench reports a 1.68 s median across four WAV turns on one laptop; this is not a live latency or accuracy guarantee.

Recorded voice bench ↗ · G2 validation plan ↗

Where the evidence comes from

Context, measurements, training candidates and validation references play different roles. A NASA source does not mean NASA trained or validated our model.

NASA / CONTEXT

NASA Human Research Program

Official hazards and risks guide the vocabulary and indicator selection. These pages are context, not physiological training datasets.

hrp-risks · hrp-hazards
NASA / MEASUREMENTS

OSDR · Inspiration4

OSD-569 and OSD-575 contain blood-test data, including glucose. Used in the project’s data analysis and context, not to train webcam drowsiness.

osdr-osd-575 · osdr-osd-569
NASA / INCIDENT REPORTS

ASRS fatigue reports

Aviation incident narratives inform the problem story. They are not astronaut incidents and do not establish incident frequency or prove the fictional film’s causal chain.

asrs-report-sets
NON-NASA / PLANNED TRAINING

UTA-RLDD + DROZY

DROZY was used in documented MediaPipe tracker probes (Massoz et al., 2016, WACV). RLDD remains planned for G2 training; no trained G2 classifier artifact is present.

uta-rldd · drozy
NON-NASA / EXPLORATORY PROBE

MRL Eye + YawDD

MRL Eye is documented in an exploratory Gemma eye-state probe. YawDD is an optional yawn-detection reference. Neither establishes spaceflight validity.

mrl-eye · yawdd
PAPERS / MEASUREMENT DESIGN

Sleep · PVT-B · PERCLOS

Astronaut sleep research motivates the sleep context; Basner’s PVT-B and the FMCSA PERCLOS brief inform measurement design. Team alert thresholds remain explicitly labeled.

jones-2022-sleep · basner-2011-pvtb · fmcsa-1998-perclos

Project source registry ↗