"Your symptoms have a pattern. We find it."

Aura is a privacy-first, local AI platform that analyzes a patient's lab results and symptom history, identifies likely autoimmune disease patterns, and produces a structured clinical report the patient can hand directly to their doctor — cutting years off the diagnostic odyssey.

Autoimmune diseases are among the most misdiagnosed and delayed-diagnosed conditions in modern medicine.

• 50+ million Americans live with an autoimmune disease
• The average patient waits 3–5 years before receiving a correct diagnosis
• They visit an average of 4–6 different doctors before anyone connects the dots
• During that window they are often told their symptoms are stress, anxiety, or "just getting older"
• Misdiagnosis leads to wrong treatments — sometimes ones that actively worsen autoimmune conditions

The diagnostic delay is not a failure of medicine. It is a failure of information flow. Each specialist sees a slice. No one sees the whole pattern.

Aura fixes that.

Aura takes everything a patient already has — their blood panels, uploaded lab PDFs, symptom descriptions, and medical photos — and runs it through a hierarchical AI pipeline that:

1. Identifies which disease cluster the patient's labs and symptoms point toward (systemic autoimmune, gastrointestinal, endocrine, or healthy)
2. Narrows to a specific likely condition within that cluster (e.g., Lupus vs. Rheumatoid Arthritis vs. Sjögren's)
3. Generates a confidence score for each finding, calibrated against 88,000+ real patient records
4. Produces two outputs — a plain-English summary the patient can understand, and a structured clinical SOAP note the doctor can act on immediately

Aura does not diagnose. It triages, patterns, and translates — giving patients the vocabulary and evidence to have a productive first conversation with the right specialist, instead of showing up with a folder of disconnected lab printouts.

• Walk into appointments with a structured, evidence-backed report rather than anecdotal descriptions
• Know which type of specialist to ask for a referral to
• Understand what their own lab values mean relative to healthy baselines and disease patterns
• End the loop of being dismissed — the report speaks the language of medicine

• Receive a pre-processed clinical summary that flags autoimmune patterns before the consultation
• Reduce time spent interpreting scattered prior records
• Make smarter referral decisions on the first visit rather than the third

• Patients arrive with the right referral and relevant prior workup already organized
• First appointment can focus on confirmation and treatment planning, not intake from scratch

• Fewer redundant specialist visits and unnecessary diagnostic panels
• Earlier correct diagnosis means treatment starts sooner — reducing long-term claims from disease progression
• One structured AI triage visit replaces 3–4 exploratory specialist visits

• Reduces unnecessary referral chains that clog specialist schedules
• Shortens the average time from first symptom to treatment plan
• Captures clinical value from lab data that would otherwise be siloed

These are conservative estimates. The compounding benefit of catching autoimmune disease 2–3 years earlier — before organ damage, disability, or treatment-resistant progression — dwarfs the direct visit savings.

Our visit simulation (notebook 07_visit_simulation.ipynb) models patient journeys across progressive clinical information — simulating what the model knows after each visit.

Key findings from 88,742 patient records:

• The model reaches >80% confidence at visit 2 for most patients
• Average 1.8 visits saved before a confident triage decision
• For systemic autoimmune diseases (Lupus, RA, Sjögren's), confident pattern detection occurs ~2 visits earlier than traditional diagnostic timelines
• At $700/visit, the average patient saves $1,260 in direct costs

For the average patient spending 3–5 years in the diagnostic odyssey, Aura compresses this to months — not by replacing doctors, but by ensuring every appointment counts.

Patient Input
    │
    ├── Lab PDFs / Blood panels
    ├── Symptom descriptions (free text)
    └── Medical photos (optional)
         │
         ▼
┌─────────────────────────────┐
│  Agent 1: Extractor         │  Parses PDFs → structured JSON
│  + Vision Model             │  Translates images → clinical keywords
└──────────────┬──────────────┘
               │
               ▼
┌─────────────────────────────┐
│  Agent 2: RAG Engine        │  Queries local PubMed vector database
│                             │  Grounds findings in peer-reviewed literature
└──────────────┬──────────────┘
               │
               ▼
┌─────────────────────────────────────────────────────┐
│  Agent 3: Hierarchical Dual-Scorer                  │
│                                                     │
│  Stage 1 — Category Classifier (XGBoost)            │
│    Trained on 88,742 patients                       │
│    Output: probability over 4 clusters              │
│      → Healthy / Systemic / GI / Endocrine          │
│    Test AUC: ~0.90                                  │
│                                                     │
│  Stage 2 — Disease Classifier (per cluster)         │
│    Systemic:      SLE, RA, Sjögren's, PsA, AS       │
│    GI:            IBD, Celiac, Functional GI        │
│    Endocrine:     Hashimoto's, Graves', T1D         │
│    Output: specific disease + confidence score      │
│    Systemic AUC: >0.92                              │
└──────────────┬──────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────┐
│  Agent 4: Translator        │  Drafts plain-English patient summary
│                             │  Drafts clinical SOAP note with DOI citations
└──────────────┬──────────────┘
               │
         ┌─────┴─────┐
         ▼           ▼
   Patient View   Doctor View
   Plain English  Clinical SOAP Note
   + Next Steps   + Literature Grounding
                  + Referral Recommendation

The model is trained on a curated, three-tier dataset architecture:

Raw lab values are transformed into clinically meaningful signals:

• Z-scores against age/sex-matched healthy baselines (not population averages)
• Inflammatory ratios — CRP/ESR ratio, neutrophil-lymphocyte ratio (NLR), platelet-lymphocyte ratio (PLR)
• Anemia pattern flags — microcytic, macrocytic, normocytic
• Autoantibody composite scores — lupus panel, RA panel, complement consumption
• Missingness flags — which labs are absent is itself a clinical signal

Aura is built local-first. No patient data leaves the device. The vector database, model inference, and report generation all run on the user's machine. This is non-negotiable for medical AI.

The UI is designed to communicate appropriate uncertainty:

• Confidence scores are always shown — never hidden
• A permanent disclaimer distinguishes pattern matching from clinical diagnosis
• Medical terms surface plain-English tooltips on hover
• The SOAP note output is clearly framed as AI-assisted, not AI-decided

aura/
├── modeling/                  # ML pipeline (see modeling/README.md)
│   ├── notebooks/             # Jupyter notebooks (numbered narrative arc)
│   ├── src/                   # Production Python modules
│   │   ├── data/              # Loaders, preprocessing, feature engineering
│   │   └── models/            # CategoryClassifier, DiseaseClassifier, DualScorer
│   ├── data/processed/        # Tiered parquet datasets
│   └── outputs/               # Figures, trained models
├── scripts/                   # Data fetch utilities
├── dataspec.md                # Full dataset specifications
└── requirements.txt

# Clone and install
git clone https://github.com/your-org/aura
pip install -r requirements.txt

# Run the modeling pipeline (in order)
jupyter lab modeling/notebooks/

# Notebooks:
# 01_data_exploration     — understand the patient population
# 02_feature_engineering  — build clinically meaningful features
# 03_baseline_models      — logistic regression baseline
# 04_advanced_models      — XGBoost, LightGBM, Random Forest, CatBoost shootout
# 05_explainability       — SHAP values, feature attribution
# 06_bias_audit           — fairness across age, sex, demographics
# 07_visit_simulation     — confidence over time, cost savings analysis

Colors: #0A0D14 background · #7B61FF violet (trust) · #3ECFCF teal (precision) · #F4A261 amber (urgency)

Typography: Clash Display headings · Inter body · JetBrains Mono scores

Animation: Three-layer background — aurora mesh gradient, Brownian particle field, mouse-tracking radial glow

Aura does not replace physicians. It gives patients the tools to be heard, and gives doctors the signal to act on.