Choose experience
Display theme
Explore platform
Designed synthetic illustration; not patient data or a diagnostic result
Assistive AI

Assistive AI with clear limits and human control

Designed to reduce administrative burden and support continuity—without independent diagnosis, prescribing, or replacement of clinical judgment. Each feature remains status-gated.

Availability to be confirmedAvailability, tier, and region require confirmation per feature
See how NAWA connectsIllustrative model; it does not represent live product data, and approved product captures are still pending.
Conceptual platform modelThe connected NAWA ecosystem
Interactive conceptual map. Visual paths illustrate two-way data flow between the NAWA center and documented platform surfaces; they do not represent live product data.
Illustrative two-way data flowIllustrative two-way flow
NAWAOne connection
Feature-specific control

One approval pattern does not fit every kind of assistance

The documented controls differ: some features create a clinician-reviewed draft, some present advisory alerts, some explain information to a patient with specific safety boundaries, and some support operations only.

Draft and review

Applies only where the feature record documents a draft or explicit approval step.

Advisory signal

Presents context to a professional and does not become a clinical decision on its own.

Explanation and limits

Patient-facing features have their own limits and escalation paths; they are not all pre-approved by a clinician.

Audience and capabilities

Choose Doctor or Patient, then inspect each capability’s status and controls

Operations is not offered because the current repository has no separately verified operations-AI category. The example cards below are illustrative, not product captures.

8 repository-documented capabilities for this audience. Every item has limited availability and an unconfirmed tier.

Voice-to-SOAP ScribeTranscribes the dictated visit audio and turns the raw transcript into a structured four-field SOAP note as a reviewable draft.Limit: Must not add facts, diagnoses, or decisions that were not supplied.
Illustrative control pattern
  1. Permitted context
  2. Produces a reviewable structured draft; the clinician remains responsible for correction and use.
  3. Explicit boundary: Must not add facts, diagnoses, or decisions that were not supplied.
Documented feature control
Produces a reviewable structured draft; the clinician remains responsible for correction and use.
Feature limitation
Must not add facts, diagnoses, or decisions that were not supplied.
Monitoring and escalation
The documented feature gate and typed failure path require runtime verification before launch.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Takes a raw transcript plus anonymized patient context and returns a SOAP note parsed from a strict JSON object with keys subjective, objective, assessment, plan.
  • Auto-detects Arabic dictation (Unicode range U+0600–U+06FF) and switches the output to clear Egyptian clinical Arabic while keeping the SOAP JSON keys in English.
  • The prompt hard-constrains the scribe — never add information, diagnoses or facts the doctor did not say — and caps generation at 800 tokens.
  • Feature-gated behind the VoiceToNote key, so it can be rate-limited or disabled independently, with failures surfaced as a typed feature exception.
SOAP Keyword ExpanderExpands the doctor's shorthand keywords into a complete professional SOAP note draft without inventing clinical facts.Limit: Must not add an unmentioned diagnosis, treatment, or dose.
Illustrative control pattern
  1. Permitted context
  2. Expands clinician shorthand into a draft for clinician review and correction.
  3. Explicit boundary: Must not add an unmentioned diagnosis, treatment, or dose.
Documented feature control
Expands clinician shorthand into a draft for clinician review and correction.
Feature limitation
Must not add an unmentioned diagnosis, treatment, or dose.
Monitoring and escalation
Input validation, output parsing, and feature-gate behavior require runtime verification.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Takes a keywords string plus patient context and returns a full SOAP note (subjective/objective/assessment/plan) via the same strict JSON parser as the voice scribe.
  • The prompt explicitly expands common Egyptian-clinic abbreviations — DM, HTN, HbA1c, met, aml — into full professional terminology.
  • Language is auto-selected by Arabic detection; Arabic output uses Egyptian clinical phrasing and preserves medication, lab and diagnosis terms verbatim.
  • A guardrail forbids adding diagnoses, treatments or doses the doctor did not mention; 800-token cap, gated behind the SoapAssist feature key.
10-Second Patient History SummaryGenerates a compact clinical history a doctor can read in about ten seconds, ending with an explicit attention-flags line.Limit: Must not infer identity or facts absent from the supplied record context.
Illustrative control pattern
  1. Permitted context
  2. Creates a read-only summary for clinician orientation; it is not an approval workflow or a source record.
  3. Explicit boundary: Must not infer identity or facts absent from the supplied record context.
Documented feature control
Creates a read-only summary for clinician orientation; it is not an approval workflow or a source record.
Feature limitation
Must not infer identity or facts absent from the supplied record context.
Monitoring and escalation
Empty-record handling and output validation are documented but need runtime evidence.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Consumes the patient context collections — conditions, allergies, medications, recent encounters, recent lab results — and returns plain text.
  • If every collection is empty it short-circuits with a 'New patient — no prior records' string and makes no AI call at all.
  • Defaults to Egyptian Arabic with an English override, enforces a 120-word limit, and closes with a fixed 'Points needing attention:' heading.
  • Output must be 3–5 short bullets, is barred from inferring name/phone/MRN/address, and is validated to reject empty text, output over 2500 chars, or any script payload; 500-token cap.
Patient-Friendly Visit SummaryTurns the finished encounter into a warm, plain-language summary for the patient — which the doctor must explicitly approve before it is saved.Limit: Must not expose raw clinical reasoning, identifiers, or unsupported instructions.
Illustrative control pattern
  1. Permitted context
  2. The documented flow requires explicit clinician approval before a patient-facing summary is saved.
  3. Explicit boundary: Must not expose raw clinical reasoning, identifiers, or unsupported instructions.
Documented feature control
The documented flow requires explicit clinician approval before a patient-facing summary is saved.
Feature limitation
Must not expose raw clinical reasoning, identifiers, or unsupported instructions.
Monitoring and escalation
Approval-state enforcement and patient delivery must be verified end to end.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Produces 150–250 words structured as greeting → what was discussed → what was found → what to do next, drawing meds only from supplied prescriptions and labs only from supplied lab orders.
  • Strips clinical detail: no raw lab values, no ICD codes, no clinical reasoning, no patient identity; uses Egyptian (not classical) Arabic when Arabic.
  • Carries an explicit approval workflow — the approve endpoint returns 409 unless the encounter is finalized, then stamps a PatientFacingSummaryApprovedAt timestamp.
  • Output is validated to 4000 chars and blocks script/javascript payloads; 700-token cap.
Clinical Decision SupportRuns a deterministic drug-allergy check first, then layers AI-detected safety concerns on top, merged and severity-ranked for the doctor.Limit: Coverage, severity, and source quality must not be treated as complete or diagnostic.
Illustrative control pattern
  1. Permitted context
  2. Presents advisory alerts for clinician assessment; alerts do not make a clinical decision.
  3. Explicit boundary: Coverage, severity, and source quality must not be treated as complete or diagnostic.
Documented feature control
Presents advisory alerts for clinician assessment; alerts do not make a clinical decision.
Feature limitation
Coverage, severity, and source quality must not be treated as complete or diagnostic.
Monitoring and escalation
Deterministic fallback, AI failure, severity ordering, and escalation behavior need clinical validation.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Layer 1 is AI-free: an in-memory allergy check cross-references patient allergies against new prescription drug families, emitting drug_allergy alerts even if the AI provider is down.
  • Layer 2 AI checks seven classes — drug-pregnancy, drug-condition contraindications, drug-drug interactions, dosage, missing lab monitoring, overdue follow-ups, duplicate therapy.
  • Severity is validated to the enum critical | warning | info (an invalid value throws), and alerts are capped at 20; if the AI fails but static alerts exist they are still returned.
  • The merge sorts critical → warning → info and, within a severity, orders the deterministic allergy_database alert ahead of the AI one; 900-token cap.
No-Show Risk ScoringScores each upcoming appointment's no-show probability with a transparent deterministic formula, then adds an AI operational insight only for higher-risk cases.Limit: Must not be used to deny care or infer protected characteristics.
Illustrative control pattern
  1. Permitted context
  2. Provides an operational risk cue for staff review; it does not automate patient action.
  3. Explicit boundary: Must not be used to deny care or infer protected characteristics.
Documented feature control
Provides an operational risk cue for staff review; it does not automate patient action.
Feature limitation
Must not be used to deny care or infer protected characteristics.
Monitoring and escalation
Bias, threshold, reason text, and intervention outcomes require monitoring before use.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Base score = no-shows ÷ total appointments, plus additive weights: +0.10 first visit, +0.15 last no-show within 30 days, +0.10 late-cancellations > 2, +0.05 for Sunday; clamped to 0–0.95.
  • Maps score to risk level (high ≥ 0.60 / medium ≥ 0.30 / low) and a suggested action (call_to_confirm ≥ 0.50 / send_reminder ≥ 0.30 / no_action), with a human reason like '2/5 prior appointments were no-shows'.
  • The AI insight is only generated when score > 0.40 — a ≤ 40-word non-judgmental note for reception, barred from names, identifiers and protected attributes; 120-token cap.
  • Processes up to 100 appointments and returns them ordered by descending probability, scores rounded to two decimals.
Smart Multi-Drug Interaction AnalysisExtends the static interaction table by having an AI pharmacology pass find pairwise and multi-drug interactions it misses, then deduplicates the combined set.Limit: Must not replace authoritative interaction sources or pharmacist and clinician judgment.
Illustrative control pattern
  1. Permitted context
  2. Adds advisory interaction analysis for clinician review alongside documented static checks.
  3. Explicit boundary: Must not replace authoritative interaction sources or pharmacist and clinician judgment.
Documented feature control
Adds advisory interaction analysis for clinician review alongside documented static checks.
Feature limitation
Must not replace authoritative interaction sources or pharmacist and clinician judgment.
Monitoring and escalation
Deduplication, severity, failure fallback, and source coverage need clinical validation.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Assembles the drug list from new prescriptions (name + dosage + frequency) plus current medications, cleaned, case-insensitively de-duplicated and capped at 30; under two drugs it returns static alerts only, no AI call.
  • The AI pass weighs pairwise and multi-drug combinations for pharmacokinetic/pharmacodynamic interactions, factoring renal impairment, liver disease, pregnancy and cumulative toxicity.
  • Returns static alerts, AI alerts (source ai_analysis) and a merged view deduplicated by normalized sorted drug set plus normalized interaction text.
  • Severity is validated to critical | warning | info, the drugs array is capped at 8 items and AI alerts at 20; 900-token cap.
Follow-Up Care PlannerAfter a completed encounter, proposes up to five prioritized, guideline-cited follow-up actions that the doctor selectively approves before anything is created.Limit: Must not autonomously create orders, referrals, or patient instructions.
Illustrative control pattern
  1. Permitted context
  2. The clinician selects and explicitly approves which draft follow-up items become actions.
  3. Explicit boundary: Must not autonomously create orders, referrals, or patient instructions.
Documented feature control
The clinician selects and explicitly approves which draft follow-up items become actions.
Feature limitation
Must not autonomously create orders, referrals, or patient instructions.
Monitoring and escalation
Approval, item creation, guideline provenance, and urgent escalation need end-to-end validation.
Tier
Tier requires product-owner confirmation

Documented technical details

  • Input includes diagnoses, new and current meds, the SOAP plan, and gap signals — last eye exam, last dental visit, last vaccination, days since last visit; output is JSON with type, title, description, timeframe, priority, reason and guideline.
  • The type enum is validated to lab_order | follow_up_visit | referral | vaccination | screening and priority to urgent | routine | when_due; the prompt asks for at most 5 recommendations (the parser hard-caps at 12), invalid JSON/enum throws.
  • Doctor approval is mandatory: the approve endpoint takes an explicit list of chosen items and rejects an empty selection with 400 — the doctor picks which drafts become real.
  • Covers standard-of-care follow-ups, medication lab-monitoring and overdue screening/referrals/vaccinations, urgent-first; 1500-token cap.
Specialty relevance

Context follows the specialty workflow; availability is never assumed

Specialty records connect documented assistance highlights to each workflow. These links explain priority; they do not prove feature, tier, or regional availability.

Dentistry

Voice-to-SOAP turns the dental visit (chart findings, procedure notes) into a structured note you review and approve

Review specialty context

Cardiology

Voice-to-SOAP captures a cardiology consult (chest pain, ECG findings, med titration) as a structured note you approve

Review specialty context

Pediatrics

Voice-to-SOAP captures a pediatric visit (fever/cough, growth, parent history) as a note you approve

Review specialty context

Obstetrics & Gynecology

Voice-to-SOAP drafts an antenatal or gynecology visit note from dictation for you to approve

Review specialty context

Oncology

Voice-to-SOAP turns an oncology visit (cycle status, side-effect review) into a structured, doctor-approved note

Review specialty context

Reproductive Medicine

Voice-to-SOAP turns a fertility consult (cycle review, treatment-cycle status) into a structured, doctor-approved note

Review specialty context
All specialties
What AI does not do

No independent clinical authority, blanket availability, or unsupported metrics

Assistance does not independently diagnose or prescribe, replace a professional, or function as emergency care. Every feature needs separate verification before an availability badge can change.

01

No autonomous clinical decision

02

No assumed tier or region

03

No accuracy or outcome claim without evidence

04

No emergency substitute

Data, monitoring, and escalation

Explicit verification gates instead of broad promises

This repository alone does not prove the model provider, hosting, retention, residency, or operational monitoring for every feature. Those details remain documented launch gates, while each capability exposes its own control and escalation record.

Data minimization

Inputs, retention, provider, and access must be approved per feature before launch.

Monitoring

Output quality, failures, bias, and boundary behavior need approved owners, measures, and response.

Escalation

There is no universal escalation promise; each feature path, channel, and response owner must be verified independently.

Review trust boundaries

Discuss a specific capability in your specialty context

The contact form carries business context only, never symptoms or patient data, and remains disabled until an approved destination is configured.

AI contact options
Useful assistance with visible human control | NAWA Digital Health