Cardio 2.013 inputs
Interactive architecture map
Autoplay activeSee which intelligence handles the request.
Follow a message from natural-language intent to the model best suited to handle it, with every role and input contract visible.
USER / NATURAL LANGUAGEperson
User messageWhat happens behind the interface?
Architecture playbackPreparing metabolic route
AUTO / READYchat_bubbleMessage
psychology
Route proposalHealthAI Reasoning ModelIntent + candidate tools
Glyco 2.016 inputs
Renal 2.024 inputs
Hepatic 2.010 inputs
PulmoVision 1.0X-ray gate
Clinical intakeReasoning path
- Detected signals
- —
- Proposed behavior
- —
- Selected intelligence
- —
- Expected input contract
- —
Why this path
The architecture will explain why each message reaches a specialist model or the clinical-intake reasoning path.
Controlled agent loop
Inspect the execution architecture.
Choose a stage to inspect its real authority boundary. Live simulator results illuminate the same path above.
Typed Pydantic contractsAllowlisted toolsONNX provenanceFail-closed fallback
security
Stage 01 · deterministic policyEmergency language pre-empts every model.
This rule engine executes before Qwen, symptom protocols, or ONNX inference. A match produces an escalation response and later stages are bypassed.
- Input
- Normalized user text
- Authority
- Deterministic rules
- Failure mode
- Bypass all models
Allowlisted tools
Small models. Explicit boundaries.
cardiologyChecking
HealthAI Cardio 2.0
Tabular screening baseline with 13 required clinical inputs.
water_dropChecking
HealthAI Glyco 2.0
Symptom questionnaire baseline with 16 explicit fields.
nephrologyChecking
HealthAI Renal 2.0
Clinical and laboratory baseline with 24 reviewed inputs.
gastroenterologyChecking
HealthAI Hepatic 2.0
Indian liver-patient baseline with 10 reviewed laboratory inputs.
clinical_notesCallable
Stateful clinical-intake reasoning
Qwen selects reviewed pathways while confirmed measurements, timing, symptoms and risk context feed an auditable disposition policy.
self_improvementCallable
Wellness guidance
General education for sleep, hydration, activity and nutrition without diagnostic claims.
radiologyExperimental
HealthAI PulmoVision 1.0
Reproducible pediatric PneumoniaMNIST baseline with a reviewed image-upload gate.
dermatologyEvaluation gate
Skin-image intake
Visible in the registry but blocked until diverse licensed data and subgroup evaluation pass.
Ready when you are
Turn your concern into a structured assessment.
Start with natural language or voice. The safety gate will guide the next supported step.
Clinical reality demo
Safety gate readyLive assessment
Real-time intake and controlled specialist-tool orchestration.
shieldSafety gate
pendingExtraction
pendingOrchestration
pendingVerification
graphic_eq
HealthAI medical voice agent Safety layer active · local inference
shield_lock Saved in this browser Medical voice AI
Severe or sudden symptoms: call India emergency services at 112.
data_object
Only confirmed state is displayedSession trace
Open-weight model playground
Meet the intelligence
Meet the intelligence
behind HealthAI.
Explore every model, follow its evidence path and run the same bounded research workflows used by the live agent. No external AI API sits behind the experience.
Checking local runtime
7Model experiences3Inference modalities0External AI APIs
neurologyHealthAIpolicy core
Choose a model
Seven models. One controlled system.
Each model has one job, an explicit input boundary and a visible evaluation status.
Selecting model workflow…
Private inference
Prompts, audio and measurements stay inside the controlled runtime.
Specialists over one giant model
Language, speech and numerical screening remain independently testable.
Evaluation before activation
A model is callable only when its artifact, schema and limitations are registered.
Evidence before activation
Every model earns its place.
Inspect frozen-test measurements, dataset lineage, runtime identity and the gates that prevent an experiment from silently becoming a clinical claim.
Loading registry
evidence_manifest.jsonpromotion_gate
artifact + schema + metrics + limitationsrequiredquery_stats
One held-out split · seed 44 · threshold selected on validation onlyFrozen-test results
These numbers reproduce the committed artifacts. They measure benchmark discrimination—not prospective clinical performance. AUROC and sensitivity must be read alongside specificity and cohort limitations.
healthai-cardio-v2.0.0
Specificity reviewHealthAI Cardio 2.0
AUROC0.8792Sensitivity1.0000Specificity0.3667Test n54
UCI Statlog Heart · 270 records · 13 structured inputs.
healthai-glyco-v2.0.0
ResearchHealthAI Glyco 2.0
AUROC0.9965Sensitivity0.9375Specificity0.9750Test n104
UCI Early Stage Diabetes · 520 records · 16 questionnaire inputs.
healthai-renal-v2.0.0
Leakage reviewHealthAI Renal 2.0
AUROC1.0000Sensitivity0.9000Specificity1.0000Test n80
UCI Chronic Kidney Disease · 400 records · 24 clinical inputs.
healthai-hepatic-v2.0.0
Calibration reviewHealthAI Hepatic 2.0
AUROC0.7548Sensitivity0.8675Specificity0.3824Test n117
UCI Indian Liver Patient · 583 records · 10 laboratory inputs.
healthai-pulmovision-v1.0.0
ExperimentalHealthAI PulmoVision 1.0
AUROC0.9267Sensitivity0.9359Specificity0.7949Test n624
PneumoniaMNIST v2 · 5,856 pediatric 28×28 benchmark images.
account_tree
Training decisions never use the frozen test splitEvaluation protocol
memory
Public name, technical release and runtime statusModel registry
| Model | Architecture | Role | Runtime | Status |
|---|---|---|---|---|
| HealthAI Cardio 2.0 | healthai-cardio-v2.0.0 | Risk pattern | Lambda CPU | Callable |
| HealthAI Glyco 2.0 | healthai-glyco-v2.0.0 | Symptom pattern | Lambda CPU | Callable |
| HealthAI Voice 1.0 | Moonshine · 34M | English voice | ARM64 Lambda | Image verified |
| HealthAI Reasoner 1.0 | Qwen3-0.6B Q8 | Extraction & routing | Planned Lambda | Not deployed |
experiment
Visible does not mean callablePromotion queue
HealthAI Reasoner 1.1
MLX LoRA adapter for evidence-bound JSON. Promotion requires held-out routing accuracy above the hybrid baseline.
HealthAI Neuro 0.1
Retired stroke artifact lacks defensible imbalance handling, calibration and complete provenance.
Multilingual Voice
Language-specific ASR remains blocked until word-error-rate and medical-term evaluations exist per language.
Research methodology01Extraction over inventionSchema-constrained proposals preserve unknown values as unknown. 02Specialists over one giant modelNumerical, language, speech and imaging artifacts remain independently testable. 03Provenance travels with outputVersion, dataset, threshold and limitations accompany every result. 04Fail closedMissing artifacts, malformed fields or blocked capabilities cannot execute.
Open weights, bounded authority.
The language model can structure, ask and propose. It cannot silently invent measurements, execute an unregistered tool or convert a benchmark score into a diagnosis.
Safety architectureModels propose.
Models propose.
Policy decides.
HealthAI is an educational research system—not a medical device, diagnosis service or treatment recommendation. Safety-critical authority stays in deterministic code.
FAIL_CLOSED = TRUEUrgent languagePre-empt every model
Priority 0Tool executionRegistry + schema required
AllowlistConversation stateEncrypted browser-held token
7-day TTLAnonymous abuseBrowser + IP + global quotas
3 layersroute
Execution order cannot be rearranged by the language modelDecision authority
shield
Four enforceable boundaries
Emergency pre-emption
Severe chest pain, breathing difficulty, facial drooping, arm weakness or loss of consciousness bypass all models. In India, call emergency services at 112.
Bounded reasoning
The reasoning model may ask, structure and propose. It cannot diagnose, invent missing measurements or execute outside the tool registry.
Confirmation before inference
Voice transcripts and structured measurements remain visible and editable before a specialist model receives them.
Minimal server knowledge
Chat history stays in localStorage. Server workflow state is encrypted and browser-held; quota storage contains only hashed counters and TTL values.
database
Data lifecycle
network_check
Abuse controls
Shared daily capacity ceilings protect model, voice, image and report infrastructure from distributed anonymous consumption.
warning
Known limitations
- Heart, diabetes, kidney and liver routes are research classifiers; none is clinically validated.
- Small public benchmarks can contain source artifacts, imbalance and population shift. Heart and liver specificity are currently weak.
- Perfect renal benchmark discrimination is a warning to investigate leakage and transportability—not evidence of clinical perfection.
- PulmoVision uses pediatric 28×28 PneumoniaMNIST images and is not a general chest-radiology model.
- HealthAI Voice 1.0 is English-only; transcripts can be wrong and require user review.
Support & collaborationschedule emergency
Talk to the team behind HealthAI.
Report a problem, suggest an evaluation case, discuss the open-model architecture, or ask about research collaboration.
alternate_emailDirect supportSauravMukherjee928@gmail.comAsynchronous support
Email is the current support channel; this is not monitored for urgent medical concerns.
Not clinical support
For severe or sudden symptoms, call India emergency services at 112.