Recalla turns short voice check-ins into longitudinal cognitive biomarker signals across speech, language, memory, and engagement, producing an illustrative CARE Score and a personal-baseline trend for caregiver and clinician decision-support.
This repository is a public technical preview: it demonstrates the engine's architecture, scientific framing, and engineering discipline on synthetic data only. It is prototype-stage, not clinically validated, and not for diagnosis.
This is a preview, not the product. Recalla's production feature extractors, scoring parameters, check-in task protocol, and longitudinal dataset are proprietary and intentionally excluded. We state that boundary openly — see
docs/whats_public_vs_private.md.
- Longitudinal first — change against a personal baseline, not a
one-time score. (
baseline.py) - Interpretable by construction — per-family sub-scores; any composite decomposes into the families that produced it.
- Reproducible — every result is stamped with engine and feature-set
versions so historical data can be re-scored.
(
registry.py) - Pluggable extraction — feature extraction sits behind a stable
interface; production extractors implement it without changing the
pipeline. (
extractors/base.py) - Human-in-the-loop — outputs prompt review, never automated decisions.
Full rationale: docs/engine_design_principles.md.
Includes: the biomarker framework, the extractor plugin interface with illustrative extractors, within-person baseline statistics, illustrative scoring and trend logic, a synthetic dataset, a runnable demo, versioning / provenance, research grounding, and accountability docs.
Excludes (proprietary): production feature extractors, fitted weights
and thresholds, the check-in task protocol, and all real data. See
docs/whats_public_vs_private.md.
Voice check-in → Feature extractors (pluggable) → Biomarker families
→ CARE Score → Personal baseline & trend → Caregiver / clinician review
Diagram: assets/care_engine_architecture.svg.
The biomarker families build on an established public literature on speech,
language, and cognition (connected-speech markers, eGeMAPS acoustic
features, DementiaBank / ADReSS benchmarks). See
docs/research_foundations.md. Consistent
with that literature, Recalla makes no diagnostic claims.
pip install -e ".[dev]"
python demo/care_score_demo.py
pytest -qThe demo scores a synthetic cohort, classifies each person's trend against
their own baseline, prints a personal-baseline deviation and a version
stamp, and writes demo/sample_output.json.
recalla-cognitive-biomarker-framework/
├── src/recalla_care_engine/
│ ├── extractors/ # pluggable feature-extraction interface + illustrative impls
│ ├── baseline.py # within-person baseline statistics
│ ├── scoring.py # illustrative sub-scores + CARE Score
│ ├── trend.py # illustrative longitudinal trend classification
│ ├── registry.py # engine / feature-set versioning + provenance
│ ├── report.py # non-diagnostic, human-readable summaries
│ └── schema.py # typed result structures
├── docs/ # framework, research, IP boundary, design, validation, privacy
├── data/ # synthetic check-ins (no real data)
├── demo/ # runnable demo + saved sample output
├── tests/ # pytest suite (CI across 3.10–3.12)
└── assets/ # architecture diagram
For demonstration, evaluation, and technical review only. Not a diagnostic
system; not for clinical decision-making; human review required. See
DISCLAIMER.md and
docs/responsible_use.md.
We are happy to walk qualified reviewers through the proprietary components under NDA — the task protocol, production feature approach, and data governance. Contact: recalla@vrealityai.com.
Prototype-stage public technical preview (v0.2.0). Clinical validation is a
future milestone — see docs/validation_roadmap.md.