Docs / Raw

Open Questions

Sourced from docs/client-implementation/OPEN_QUESTIONS.md

Edit on GitHub

VoiceAssist Client Development - Open Questions & Decisions

Version: 2.0 Date: 2025-11-21 Status: ✅ Critical Decisions Made - Ready for Milestone 1 Branch: client-roadmap-reconciliation


🎉 Critical Decisions Made (2025-11-21)

All 8 critical questions have been resolved! Development can proceed.

QuestionDecisionImpact
Q1: Design System✅ Create from scratch (Radix UI + Tailwind)Week 1-2
Q2: Storybook✅ Yes, include in monorepo setupWeek 1-2
Q6: Deployment✅ Ubuntu server (Docker Compose) initiallyWeek 1-2
Q10: UpToDate License✅ No budget, use free sourcesMilestone 5
Q15: Offline PHI✅ No PHI offline, non-PHI onlyMilestone 6
Q18: GPU Budget✅ No budget, use OpenAI APIsMilestone 3
Q22: Image Datasets✅ Use pre-trained models (GPT-4 Vision)Milestone 6
Q23: AI Liability✅ Decision support only, clear disclaimersAll phases

See Resolved Decisions section for details.


Purpose

This document consolidates all open questions that require answers before proceeding with client development. Questions are organized by category and priority.

Total Questions: 23

  • Critical (resolved): 8 questions ✅
  • Medium Priority (answer by specific milestones): 10 questions
  • Low Priority (can decide later): 5 questions

Summary Dashboard

CategoryTotalCriticalMediumLow
Design & UX5221
Infrastructure & Operations4220
External Dependencies5131
Compliance & Security3120
AI & Machine Learning6213
Total238105

Table of Contents

  1. Critical Decisions
  2. Medium Priority Decisions
  3. Low Priority Decisions
  4. Decision Template
  5. Approval Process

Resolved Decisions

Critical Questions - Final Decisions (2025-11-21)

All 8 critical questions have been answered and documented below. These decisions are final and development can proceed.


✅ Q1: Design System Availability - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: Create design system from scratch

Rationale:

  • No existing design system available
  • Create basic design system using Radix UI + Tailwind as foundation
  • Focus on medical professionalism and trust-building design

Implementation Details:

  • Design Tokens Package: @voiceassist/design-tokens

    • Colors: Medical blues, teals, grays (trust-building palette)
    • Typography: System fonts (San Francisco, Segoe UI, Roboto)
    • Spacing: 4px/8px grid system
    • Border radius, shadows, z-index scales
    • Animation timing and easing functions
    • Breakpoints for responsive design
  • Component Library: @voiceassist/ui

    • Base: Radix UI primitives + Tailwind CSS
    • Documentation: Storybook
    • Accessibility: WCAG 2.1 AA compliance from day 1

Timeline: Week 1-2 (Phase 0) Status: Ready to implement


✅ Q2: Storybook Setup - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: Yes, include Storybook in monorepo setup

Rationale:

  • Component documentation from day 1
  • Visual testing during development
  • Accessibility testing integration (axe-core)
  • Better collaboration and design system showcase

Implementation Details:

  • Setup: Storybook 8.0+ with Vite

  • Addons:

    • Accessibility addon (axe-core)
    • Docs addon for MDX documentation
    • Controls addon for interactive props
    • Viewport addon for responsive testing
    • Interactions addon for testing
  • Deployment: Local development only initially

  • Timeline: Week 1 (during monorepo setup)

Status: Ready to implement


✅ Q6: Deployment Strategy - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: Ubuntu server (Docker Compose) initially, evaluate Vercel/Netlify after Milestone 1

Rationale:

  • Leverage existing production-ready Ubuntu server infrastructure
  • Simple deployment via Docker Compose (same as backend)
  • No additional cost or complexity
  • Re-evaluate managed frontend hosting (Vercel/Netlify) after Milestone 1 for:
    • Global CDN
    • Automatic SSL
    • Preview deployments

Implementation Details:

  • Backend: Ubuntu server (asimo.io) - existing

  • Frontend: Ubuntu server (Docker Compose) for Milestone 1

    • Deploy web-app, admin-panel, docs-site as Docker containers
    • Nginx reverse proxy (existing)
    • Let's Encrypt SSL (existing)
  • Subdomains:

    • voiceassist.asimo.io → web-app
    • admin.voiceassist.asimo.io → admin-panel
    • docs.voiceassist.asimo.io → docs-site
  • Future Evaluation (after Milestone 1):

    • Move frontend to Vercel/Netlify for better CDN
    • Keep backend on Ubuntu server
    • Benefits: Global CDN, auto-scaling, preview deployments

Timeline: Week 1-2 (Docker Compose), Re-evaluate Week 10 Status: Ready to implement


✅ Q10: UpToDate Licensing Budget - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: No budget allocated, use free sources

Rationale:

  • UpToDate license ($500-1000/month) not in current budget

  • Excellent free alternatives available:

    • PubMed (35M+ citations, free)
    • OpenEvidence (evidence synthesis, free tier)
    • Clinical practice guidelines (CDC, WHO, free)
  • Re-evaluate in later milestones if user demand justifies cost

Implementation Details:

  • Primary Sources (Milestone 5):

    1. PubMed - Week 40-42 (highest priority)
    2. OpenEvidence - Week 37-38 (high priority)
    3. Clinical Guidelines - Week 39 (CDC, WHO, specialty societies)
  • Alternative (if budget later):

    • DynaMed ($400-800/month) as UpToDate alternative

Timeline: Milestone 5 (Weeks 37-44) Status: Documented, proceed with free sources


✅ Q15: Offline Mode PHI Regulations - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team (with compliance consideration) Final Decision: No PHI in offline storage, non-PHI only with encryption

Rationale:

  • HIPAA compliance: Safest approach is no offline PHI
  • Minimize risk of PHI exposure on lost/stolen devices
  • Offline functionality limited to non-PHI features

Implementation Details:

  • Allowed Offline (non-PHI):

    • Medical knowledge base articles
    • De-identified conversation history (remove names, dates, identifiers)
    • User preferences and settings (theme, language)
    • UI assets (JavaScript, CSS, images)
  • Prohibited Offline (PHI):

    • Patient demographics
    • Clinical context (problems, medications, labs)
    • Identifiable data
    • Actual patient information
  • Security Measures:

    • AES-256 encryption for all offline data
    • Auto-expiration after 7 days
    • Re-authentication required to access cached data
    • Secure deletion on logout
  • User Consent:

    • Prompt on first offline mode use
    • Clear explanation of what's cached
    • Option to disable offline mode

Timeline: Milestone 6 (Weeks 45-52) Status: Policy documented, ready to implement


✅ Q18: GPU Infrastructure Budget - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: No budget for GPU infrastructure, use hosted OpenAI APIs

Rationale:

  • GPU infrastructure ($500-1500/month) not in current budget
  • OpenAI APIs provide good medical performance
  • Hosted services (OpenAI) sufficient for MVP
  • Re-assess GPU needs after core system is live and user feedback collected

Implementation Details:

  • Current Approach (Milestone 3 and beyond):

    • Continue with OpenAI text-embedding-3-small for embeddings
    • Use OpenAI GPT-4 for generation
    • Monitor costs and accuracy
  • Future Re-evaluation (Month 9+):

    • If OpenAI costs exceed $500/month: Consider GPU
    • If medical accuracy < 85%: Consider BioGPT/PubMedBERT
    • If data sovereignty becomes important: Consider self-hosted models
  • Alternative Hosted Options:

    • Hugging Face Inference API ($200-600/month, managed)
    • AWS SageMaker Serverless (pay per use)

Timeline: Milestone 3 (Weeks 21-28), Re-evaluate Month 9 Status: Documented, proceed with OpenAI APIs


✅ Q22: Medical Image Datasets - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: Use pre-trained models (GPT-4 Vision), evaluate public datasets later

Rationale:

  • Pre-trained GPT-4 Vision provides good accuracy (80-90%) out-of-the-box
  • No need for custom training or labeled datasets initially
  • Public datasets available if needed later (HAM10000, ChestX-ray14)

Implementation Details:

  • Initial Approach (Milestone 6):

    • GPT-4 Vision for medical image analysis
    • Support use cases:
      1. Dermatology (skin lesions)
      2. Wound assessment
      3. General medical images
  • Accuracy Validation:

    • Benchmark on test dataset
    • Medical professional review
    • User feedback collection
  • Future Options (if accuracy < 85%):

    • Free Datasets:

      • HAM10000 (dermatology, 10k images, free)
      • ChestX-ray14 (radiology, 100k images, free)
      • RSNA datasets (various competitions, free with attribution)
    • Fine-tuning: If dataset acquired and accuracy improvement > 10%

Timeline: Milestone 6 (Weeks 45-52) Status: Documented, proceed with GPT-4 Vision


✅ Q23: AI Diagnosis Liability - RESOLVED

Decision Date: 2025-11-21 Decision Maker: Product Team Final Decision: Decision support tool only, clear disclaimers, no diagnostic claims

Rationale:

  • Avoid FDA medical device classification
  • Lower liability risk
  • Faster to market (no FDA approval required)
  • Clear positioning as educational/decision support tool

Implementation Details:

  • Positioning:

    • "Clinical decision support tool"
    • "For educational and reference purposes"
    • "Not a substitute for professional medical judgment"
    • "Not for diagnostic use"
  • Required Disclaimers:

    1. On First Use:

      VoiceAssist is a clinical decision support tool, not a diagnostic device.
      All AI-generated information should be verified with primary sources and
      should not replace professional medical judgment. By using this tool,
      you acknowledge that VoiceAssist is for educational and reference
      purposes only.
      
      • User must accept terms before using
    2. Footer on Every Page:

      • "Decision support tool. Not for diagnostic use."
    3. On AI Responses (especially image analysis):

      • "Possible findings: [description]. Recommend [action]. Verify with primary sources."
      • Always cite sources
      • Encourage verification
  • Prohibited Language:

    • ❌ "Diagnosis: [condition]"
    • ❌ "VoiceAssist diagnoses..."
    • ❌ "Diagnostic accuracy of 95%"
    • ✅ "Possible findings suggest..."
    • ✅ "Consider differential diagnosis including..."
    • ✅ "Recommend consulting [specialist]..."
  • Legal Requirements:

    • Terms of Service updated with disclaimers
    • User acceptance flow on first use
    • Professional liability insurance review
    • Medical advisory board review of all AI features

Timeline: All phases (implement disclaimers in Week 3) Status: Policy documented, ready to implement


Critical Decisions (Needed Before Starting)

Design & UX

Q1: Design System Availability ⚠️ CRITICAL

Category: Design & UX Priority: P0 (Critical) Impact: HIGH - Affects Week 1-2 timeline Decision Needed By: Before Milestone 1 starts (Week 1)

Question: Does a design system already exist (Figma/Sketch files, style guide, component library)?

Options:

  • A) Existing design system available

    • Use as-is, create design tokens from it
    • Timeline: 1 week
  • B) Partial design system exists

    • Complete missing pieces (colors, typography, components)
    • Timeline: 1.5 weeks
  • C) No design system exists

    • Create from scratch using medical UI best practices
    • Timeline: 2 weeks

Provisional Answer: Assume Option C (no existing design system).

Recommended Approach:

  1. Research medical UI references:

    • Medscape, UpToDate, Epic MyChart
    • Healthcare.gov, Patient portals
  2. Create professional medical design system:

    • Color palette: Medical blues, teals, grays (trust-building)
    • Typography: System fonts (San Francisco, Segoe UI, Roboto) for professionalism
    • Spacing: 4px/8px grid system
    • Components: Based on shadcn/ui + Radix UI
  3. Document in Figma + Storybook

  4. Review with medical professionals for feedback

Timeline: 2 weeks (Phase 0) Effort: 1 designer + 1 developer

What We Need:

  • Confirmation of existing design system availability
  • Brand guidelines (if any)
  • Logo files (SVG, PNG)
  • Color preferences
  • Typography preferences

Impact of Delay: If this decision is delayed, Week 1-2 timeline extends to Week 1-3.


Q2: Storybook Setup ⚠️ CRITICAL

Category: Design & UX Priority: P0 (Critical) Impact: MEDIUM - Affects Week 1-2 tasks Decision Needed By: Week 1

Question: Should Storybook be part of the initial monorepo setup?

Options:

  • A) Yes, set up Storybook in Week 1-2

    • Better component documentation from day 1
    • Visual testing during development
    • Accessibility testing integration (axe-core)
    • Easier collaboration with designers
    • Effort: +0.5 week
  • B) Defer to Week 10+

    • Faster initial setup
    • Add documentation later
    • Effort: 1 week later

Provisional Answer: Recommend Option A (set up Storybook in Week 1-2).

Benefits:

  • Component documentation from day 1
  • Visual regression testing
  • Accessibility testing with axe-core
  • Designer collaboration (share Storybook URL)
  • Isolated component development
  • Design system showcase

Drawbacks:

  • Adds 0.5 week to Phase 0
  • Requires Storybook maintenance

Recommended Setup:

  • Storybook 7.0+ with Vite
  • Accessibility addon (axe-core)
  • Docs addon for MDX documentation
  • Controls addon for interactive props
  • Viewport addon for responsive testing

What We Need:

  • Approval to extend Phase 0 by 0.5 week if needed
  • Decision: Deploy Storybook or keep local only?

Impact of Delay: If deferred to Week 10+, component documentation will be incomplete, making collaboration harder.


Infrastructure & Operations

Q6: Deployment Strategy ⚠️ CRITICAL

Category: Infrastructure Priority: P0 (Critical) Impact: HIGH - Affects deployment architecture Decision Needed By: Week 1-2

Question: Should frontend apps deploy to the same Ubuntu server or separate infrastructure?

Options:

  • A) Same Ubuntu server (asimo.io)

    • Pros: Simpler, single point of management, no extra cost
    • Cons: Single point of failure, no global CDN, manual deployments
    • Cost: $0 extra
  • B) Separate infrastructure (Vercel/Netlify for frontend)

    • Pros:
      • Global CDN (faster load times worldwide)
      • Automatic SSL and preview deployments
      • Frontend updates independent of backend
      • Free tier sufficient for testing
      • Auto-scaling
    • Cons: Additional service to manage
    • Cost: $0-20/month (free tier likely sufficient)
  • C) Kubernetes cluster for all

    • Pros: Most scalable, enterprise-grade
    • Cons: Most complex, highest cost
    • Cost: $100-300/month

Provisional Answer: Recommend Option B (hybrid approach):

Architecture:

  • Backend: Ubuntu server at asimo.io (existing, production-ready)

    • API Gateway, databases, services
    • Existing monitoring stack
  • Frontend: Vercel or Netlify

    • voiceassist.asimo.io → Vercel/Netlify edge
    • admin.voiceassist.asimo.io → Vercel/Netlify edge
    • docs.voiceassist.asimo.io → Vercel/Netlify edge
    • Global CDN for fast load times
    • Automatic preview deployments for PRs

Benefits:

  • Frontend updates don't require backend deployment
  • Global CDN improves performance
  • Free tier sufficient for development/testing
  • Preview deployments for easy QA
  • Automatic SSL certificates

Configuration:

  • DNS: Point frontend subdomains to Vercel/Netlify
  • CORS: Update backend to allow frontend origins
  • Environment variables: API_URL, WebSocket URL

Cost Estimate:

  • Vercel/Netlify: $0-20/month (free tier: 100GB bandwidth)
  • Total: $0-20/month extra

What We Need:

  • Approval for Vercel/Netlify usage
  • DNS access to configure subdomains
  • Approval for $0-20/month budget (if exceeding free tier)

Alternative (if Option B rejected): Deploy to Ubuntu server with Nginx serving static files and reverse proxy to backend.


Q18: GPU Infrastructure Budget ⚠️ CRITICAL

Category: AI & Machine Learning Priority: P0 (Critical) Impact: HIGH - Affects Milestone 3 (Advanced AI) Decision Needed By: Week 20 (before Milestone 3)

Question: Do we have budget/resources for GPU infrastructure to run BioGPT/PubMedBERT?

Background: Phase 5 (Medical AI) was completed with OpenAI embeddings (MVP). Specialized medical models (BioGPT, PubMedBERT) were deferred due to GPU infrastructure requirements.

Options:

  • A) Yes, budget approved for GPU infrastructure

    • Benefits:
      • Medical-specific embeddings (higher accuracy)
      • Lower cost per query (vs OpenAI long-term)
      • Data sovereignty (no data sent to OpenAI)
      • Fine-tuning possible
    • Cost: $500-1500/month
    • Options:
      • AWS EC2 g4dn.xlarge: $500/month (dedicated)
      • AWS SageMaker inference: $300-800/month (pay per use)
      • Hugging Face Inference API: $200-600/month (managed)
  • B) No budget for GPU infrastructure

    • Impact: Continue with OpenAI embeddings
    • Cost: $0 extra (included in OpenAI API costs)
    • Accuracy: Good but not medical-optimized

Provisional Answer: Recommend Option A if budget allows, with cost-benefit analysis:

Year 1 Projection:

  • OpenAI embeddings: ~$300-500/month (at 100k queries/month)
  • BioGPT self-hosted: $500/month (GPU) + minimal API costs
  • Break-even: ~6 months
  • Year 1 savings: $1,200-2,400

Accuracy Improvement (estimated):

  • RAG precision: +5-10%
  • Medical entity recognition: +15-20%
  • Domain-specific queries: +20-30%

Recommended Approach:

  1. Month 1-3: Use OpenAI embeddings (current)
  2. Month 4: Evaluate BioGPT/PubMedBERT on test dataset
  3. Month 5: If accuracy improvement > 10%, migrate
  4. Month 6+: Self-hosted medical models

What We Need:

  • Budget approval for $500-1500/month GPU infrastructure
  • Choice of cloud provider (AWS, GCP, Azure)
  • IT approval for GPU instance provisioning

Impact if No Budget: Continue with OpenAI embeddings. Accuracy will be good but not medical-optimized. No data sovereignty.


Q10: UpToDate Licensing Budget ⚠️ CRITICAL

Category: External Dependencies Priority: P0 (Critical) Impact: HIGH - Affects Milestone 5 Decision Needed By: Before Milestone 5 (Week 37)

Question: What's the budget for UpToDate licensing (~$500-1000/month)?

Background: UpToDate is the gold-standard clinical decision support tool used by healthcare professionals worldwide. Integration requires a commercial license.

Options:

  • A) Budget approved for UpToDate

    • Cost: $500-1000/month (~$6,000-12,000/year)
    • Benefits:
      • 11,500+ clinical topics
      • Drug interaction database
      • Diagnostic algorithms
      • Evidence-based recommendations
      • Trusted by 2M+ clinicians worldwide
  • B) No budget for UpToDate

    • Cost: $0
    • Impact: Focus on free sources
    • Alternatives:
      • PubMed (free, 35M+ citations)
      • OpenEvidence (free tier, evidence synthesis)
      • Clinical practice guidelines (free, CDC, WHO)
      • DynaMed (alternative, $400-800/month)

Provisional Answer: Budget approval needed. UpToDate is highly valuable but expensive.

Cost-Benefit Analysis:

  • Value per User: If 100 clinicians use VoiceAssist daily
    • Cost per user: $5-10/month
    • Time saved: ~30 min/day (faster lookups)
    • Value: $50-100/month per clinician
    • ROI: 5-10x

Recommended Approach:

  1. Start with free sources: PubMed, OpenEvidence, guidelines
  2. Evaluate user feedback: Do users need UpToDate?
  3. Month 6: If high demand, pursue UpToDate license
  4. Year 1: Re-evaluate based on usage metrics

Alternatives (if no budget):

  • DynaMed: $400-800/month (alternative to UpToDate)
  • ClinicalKey: $300-600/month (Elsevier)
  • PubMed + OpenEvidence: Free

What We Need:

  • Budget approval for $500-1000/month
  • Legal approval for commercial API license
  • Decision on alternatives if UpToDate not approved

Impact if No Budget: No UpToDate integration. Use free sources (PubMed, OpenEvidence). Users may need to reference UpToDate separately.


Compliance & Security

Q15: Offline Mode PHI Regulations ⚠️ CRITICAL

Category: Compliance & Security Priority: P0 (Critical) Impact: HIGH - Affects Milestone 6 (Offline/PWA) Decision Needed By: Before Milestone 6 (Week 45)

Question: What are the regulatory constraints on offline PHI storage under HIPAA?

Background: Milestone 6 includes offline mode and PWA features. Storing PHI offline (on user's device) requires careful HIPAA compliance.

Options:

  • A) PHI allowed offline with proper encryption

    • Requirements:
      • AES-256 encryption for offline data
      • Auto-expiration of offline data (7-30 days)
      • User consent for offline storage
      • Remote wipe capability
      • Audit trail for offline access
      • Business Associate Agreement (BAA) updates
    • Benefits: Full offline functionality
    • Risks: PHI exposure if device lost/stolen
  • B) PHI not allowed offline (non-PHI only)

    • Offline Storage Allowed:
      • Medical knowledge base articles
      • De-identified conversation history
      • User preferences and settings
      • Cached UI assets
    • Offline Storage Prohibited:
      • Patient demographics
      • Clinical context (problems, medications, labs)
      • Identifiable data
    • Benefits: Simpler compliance, lower risk
    • Limitations: Reduced offline functionality
  • C) Consult HIPAA compliance officer

    • Get official guidance before implementing

Provisional Answer: Strongly recommend Option C (consult compliance officer), then implement Option B (non-PHI only) for safety.

Recommended Approach:

  1. Before Week 45: Consult with HIPAA compliance officer
  2. Get written approval for offline PHI storage (if allowed)
  3. Implement Option B (non-PHI only) as default
  4. If Option A approved: Add encrypted PHI storage as optional feature

Non-PHI Offline Mode (Recommended Baseline):

  • Medical knowledge base articles (cached)
  • De-identified conversation history:
    • Remove patient names, dates, identifiers
    • Cache conversation text and citations
    • Expire after 7 days
  • User preferences (theme, language, voice settings)
  • UI assets (JavaScript, CSS, images)

PHI Offline Mode (If Approved):

  • Encrypted IndexedDB storage (AES-256)
  • Auto-expiration after 24 hours
  • Remote wipe via backend API
  • User consent prompt on first use
  • Audit log sent to backend on sync
  • Device PIN/biometric required to access

What We Need:

  • Meeting with HIPAA compliance officer
  • Written approval for offline PHI storage (or denial)
  • Updated Business Associate Agreement (if PHI offline allowed)
  • Security review of offline encryption approach

Impact if PHI Offline Not Allowed: Offline mode limited to non-PHI features. Users must be online for clinical features.

Decision Deadline: Week 40 (5 weeks before Milestone 6)


AI & Machine Learning

Q22: Medical Image Datasets ⚠️ CRITICAL

Category: AI & Machine Learning Priority: P0 (Critical) Impact: HIGH - Affects Milestone 6 (Multi-Modal AI) Decision Needed By: Before Milestone 6 (Week 45)

Question: Do we have access to labeled medical image datasets for training/fine-tuning?

Background: Milestone 6 includes multi-modal AI (medical image analysis). Custom models require labeled medical images.

Options:

  • A) Yes, licensed datasets available

    • Use Cases:
      • Train custom medical image classifiers
      • Fine-tune existing models
      • Benchmark accuracy
    • Datasets:
      • Dermatology: HAM10000 (10,000 images, free)
      • Radiology: ChestX-ray14 (100,000 images, free)
      • Radiology: MIMIC-CXR (377,000 images, license required)
    • Effort: 4-6 weeks for custom model
  • B) No, use pre-trained models only

    • Use GPT-4 Vision:
      • General medical image analysis
      • No custom training needed
      • Accuracy: good but not specialized
    • Effort: 2-3 weeks for integration
    • Cost: $0.01-0.02 per image analysis

Provisional Answer: Recommend Option B (pre-trained models) initially.

Recommended Approach:

  1. Start with GPT-4 Vision:

    • General medical image analysis
    • Supports dermatology, wounds, ECG, X-rays
    • No training data needed
    • Good accuracy out-of-the-box
  2. Evaluate accuracy:

    • Benchmark on test dataset
    • Get feedback from medical professionals
    • Measure: precision, recall, F1 score
  3. If accuracy < 85%:

    • Evaluate fine-tuning with labeled datasets
    • Consider domain-specific models (DermNet, CheXNet)
  4. Month 12+: Custom models if needed

Free Datasets Available:

  • HAM10000 (dermatology, 10k images, free)
  • ChestX-ray14 (radiology, 100k images, free)
  • PAD-UFES-20 (skin lesions, 2k images, free)

Licensed Datasets:

  • MIMIC-CXR (radiology, 377k images, requires PhysioNet license)
  • NIH Chest X-ray (112k images, free but citation required)

What We Need:

  • Decision: Pre-trained models only or custom training?
  • If custom: Dataset license approvals
  • If custom: GPU infrastructure (see Q18)
  • Legal review of dataset terms

Impact if No Datasets: Use GPT-4 Vision only. Accuracy will be good (80-90%) but not medical-optimized (90-95%).


Q23: AI Diagnosis Liability ⚠️ CRITICAL

Category: AI & Machine Learning Priority: P0 (Critical) Impact: CRITICAL - Legal/regulatory Decision Needed By: Before any image analysis feature (Week 40)

Question: What are the liability considerations for AI-assisted diagnosis? Should we pursue FDA approval?

Background: Medical AI systems that provide diagnostic advice may be considered medical devices requiring FDA approval.

Options:

  • A) Decision support only (no diagnosis claims)

    • Positioning:
      • "Educational and decision support tool"
      • "Not a substitute for professional medical judgment"
      • "Always verify with primary sources"
    • Disclaimers: Clear, prominent, require user acknowledgment
    • FDA: Not required for decision support
    • Liability: Lower risk
    • Effort: Legal review, disclaimers
  • B) Diagnostic assistance with disclaimers

    • Positioning:
      • "AI-assisted diagnostic support"
      • Provides differential diagnosis suggestions
      • Confidence scores shown
    • FDA: May require 510(k) clearance (medical device)
    • Liability: Medium risk
    • Effort: 6-12 months FDA approval, $50k-200k
  • C) Full diagnostic system (FDA-cleared)

    • Positioning:
      • "FDA-cleared diagnostic AI"
      • Direct diagnostic capabilities
    • FDA: Requires De Novo or PMA approval
    • Liability: Highest risk, highest value
    • Effort: 12-24 months, $200k-1M+

Provisional Answer: Strongly recommend Option A (decision support only).

Recommended Approach:

  1. Position as decision support tool:

    • "For educational purposes"
    • "Clinical decision support"
    • "Not for diagnostic use"
  2. Implement comprehensive disclaimers:

    • On first use: User must accept terms
    • On every image analysis result
    • In footer of every page
    • In Terms of Service
  3. Clear user acknowledgment:

    "VoiceAssist is a clinical decision support tool, not a diagnostic device.
    All AI-generated information should be verified with primary sources and
    should not replace professional medical judgment. By using this tool,
    you acknowledge that VoiceAssist is for educational and reference
    purposes only."
    
  4. Avoid diagnostic language:

    • ❌ "Diagnosis: Melanoma"
    • ✅ "Possible findings: Pigmented lesion with asymmetry. Suggest dermatology referral."
  5. Always cite sources:

    • Link to guidelines, literature
    • Encourage verification

Legal Requirements:

  • Legal review of all AI features
  • Terms of Service updated with disclaimers
  • User acceptance flow implemented
  • Professional liability insurance review
  • Consult with FDA regulatory expert

What We Need:

  • Legal review before Week 40
  • Decision: Decision support only or pursue FDA approval?
  • If FDA: Budget $50k-200k, timeline 6-12 months

Impact: If we pursue FDA approval: 6-12 month delay, $50k-200k cost, but much higher value and trust.

If we stay decision support: Faster to market, lower cost, but cannot make diagnostic claims.


Medium Priority Decisions (By Milestone)

Design & UX

Q3: Component Library Strategy

Priority: P2 (Medium) Impact: MEDIUM - Affects maintainability Decision Needed By: Week 1

Question: Should we use shadcn/ui as-is or fork and customize extensively?

Options:

  • A) Use shadcn/ui as-is

    • Easier updates from upstream
    • Community support
    • Customize via design tokens only
  • B) Fork and customize

    • More control over components
    • Harder to update
    • Full customization possible

Provisional Answer: Recommend Option A (use shadcn/ui as-is with theme customization).

Rationale:

  • shadcn/ui is highly customizable via Tailwind
  • Design tokens provide sufficient control
  • Forking creates maintenance burden
  • Can always fork specific components later if needed

Recommendation:

  • Start with shadcn/ui + design tokens
  • Customize colors, typography, spacing via tokens
  • Fork only if specific component needs major changes

Q4: Dark Mode Priority

Priority: P3 (Low) Impact: LOW - Can be added later Decision Needed By: Week 10

Question: Should dark mode be in MVP or deferred?

Options:

  • A) MVP (Week 2)

    • More work upfront
    • Both themes from day 1
    • Effort: +0.5 week
  • B) Defer to Week 10+

    • Faster MVP
    • Focus on light mode first
    • Add dark mode in polish phase

Provisional Answer: Recommend Option B (defer to Week 10).

Rationale:

  • Light mode sufficient for MVP
  • Medical professionals typically work in well-lit environments
  • Dark mode can be added in polish phase (Week 10)
  • Saves 0.5 week in Phase 0

Recommendation:

  • Build light mode first
  • Design tokens support dark mode (prepare colors)
  • Implement dark mode in Week 10 (Phase 2, Advanced Features)

Infrastructure & Operations

Q7: Staging Environments

Priority: P1 (High) Impact: MEDIUM - Affects testing workflow Decision Needed By: Week 2

Question: Do we need separate staging/production environments for frontend?

Options:

  • A) Yes, separate staging environments

    • Staging: staging.voiceassist.asimo.io
    • Production: voiceassist.asimo.io
    • Benefits: Safer, final QA before production
    • Cost: $0 (same Vercel/Netlify account)
  • B) No, test locally + preview deployments

    • Testing: Local dev + PR preview deployments
    • Benefits: Faster, simpler
    • Risks: No final QA environment

Provisional Answer: Recommend Option A (separate staging environments).

Configuration:

Workflow:

  1. Develop locally
  2. Open PR → preview deployment
  3. Merge to develop → deploy to staging
  4. Test on staging
  5. Merge to main → deploy to production

Cost: $0 (same Vercel/Netlify account supports multiple environments)


Q9: Telemetry Provider

Priority: P1 (High) Impact: MEDIUM - Affects cost and features Decision Needed By: Week 19 (when implementing telemetry package)

Question: Which telemetry provider for client-side errors and performance monitoring?

Options:

  • A) Sentry (error tracking)

    • Cost: $26/month (50k errors/month), $79/month (250k errors/month)
    • Features: Error tracking, performance monitoring, session replay
    • Pros: Best error tracking, affordable, good React support
  • B) DataDog (full observability)

    • Cost: $15/host/month + $5/million spans
    • Features: Logs, metrics, traces, RUM, profiling
    • Pros: Full observability suite
    • Cons: Expensive for full features
  • C) New Relic (balanced)

    • Cost: $99/month (100GB data), $349/month (unlimited)
    • Features: APM, browser monitoring, dashboards
    • Pros: Good balance of features and cost
  • D) Self-hosted (Grafana Loki + Tempo)

    • Cost: $0 (infrastructure cost only)
    • Features: Logs, traces (limited error tracking)
    • Pros: Free, data sovereignty
    • Cons: More setup, limited features

Provisional Answer: Recommend hybrid approach:

  • Client-side errors: Sentry ($26-79/month)
  • Backend observability: Existing Grafana stack (Prometheus, Loki, Jaeger)

Rationale:

  • Sentry excels at client-side error tracking
  • Grafana stack already set up for backend
  • Best of both worlds, reasonable cost

Budget Estimate:

  • Development: Sentry free tier (5k errors/month)
  • Production: Sentry $26-79/month (50k-250k errors/month)
  • Year 1: ~$300-900/year

What We Need:

  • Budget approval for $26-79/month ($300-900/year)
  • Sentry account setup
  • GDPR review (if applicable)

External Dependencies

Q11: External API Priorities

Priority: P1 (High) Impact: MEDIUM - Affects Milestone 5 sequencing Decision Needed By: Week 35

Question: What's the priority ranking for external medical integrations?

Options to Rank:

  • UpToDate (clinical decision support)
  • OpenEvidence (evidence-based medicine)
  • PubMed (literature search)
  • Clinical trial databases (ClinicalTrials.gov)
  • Drug information systems (Micromedex, Lexicomp)

Provisional Answer: Recommended Prioritization:

  1. PubMed (P1 - Highest Priority)

    • Cost: Free
    • Benefits: 35M+ citations, essential for literature search
    • Effort: 2 weeks
    • Timeline: Week 40-42
  2. OpenEvidence (P1 - High Priority)

    • Cost: Free tier available
    • Benefits: Evidence synthesis, clinical questions
    • Effort: 1 week
    • Timeline: Week 37-38
  3. UpToDate (P0 if licensed, P3 if not)

    • Cost: $500-1000/month (requires license)
    • Benefits: Best clinical decision support, 11,500+ topics
    • Effort: 2 weeks
    • Timeline: Week 37-39 (if licensed)
  4. Drug Information (P2 - Medium Priority)

    • Cost: Varies ($200-500/month)
    • Benefits: Drug interactions, dosing, safety
    • Effort: 1-2 weeks
    • Timeline: Milestone 7+ (future)
  5. Clinical Trials (P3 - Low Priority)

    • Cost: Free (ClinicalTrials.gov API)
    • Benefits: Trial matching, enrollment
    • Effort: 1 week
    • Timeline: Milestone 8+ (future)

Rationale:

  • PubMed is free and essential for citations
  • OpenEvidence provides evidence synthesis for free
  • UpToDate requires budget approval
  • Drug info and trials can be deferred

What We Need:

  • Confirmation of priority ranking
  • Budget approval for UpToDate (if high priority)

Q12: EMR Integration Targets

Priority: P2 (Medium) Impact: HIGH - Affects future EMR integration Decision Needed By: Month 6

Question: Are there specific hospital partners or EMR systems to target first?

Options:

  • A) Specific hospital partnership in progress

    • Focus on their EMR system first
    • Customized integration
  • B) Most popular EMRs (Epic, Cerner, Allscripts)

    • Epic: 31% market share
    • Cerner: 25% market share
    • Allscripts: 12% market share
  • C) Generic FHIR R4 standard

    • Works with all FHIR-compliant EMRs
    • Start with read-only patient data
    • Expand to read-write later

Provisional Answer: Recommend Option C (generic FHIR R4) initially.

Rationale:

  • HL7 FHIR R4 is standard across major EMRs
  • Avoid vendor lock-in
  • Easier to add hospital-specific customizations later
  • Start with read-only to reduce complexity

Recommended Approach:

  1. Implement FHIR R4 read-only:

    • Patient demographics
    • Observations (labs, vitals)
    • Conditions (diagnoses)
    • Medications
  2. Test with Epic/Cerner sandboxes:

    • Both provide free developer sandboxes
    • Validate FHIR compliance
  3. Expand based on partnerships:

    • Add hospital-specific features as needed
  4. Timeline: Milestone 7+ (Month 13+)

What We Need:

  • Hospital partnership status
  • Preferred EMR systems
  • Read-only vs read-write requirements

Q14: FHIR Certification

Priority: P2 (Medium) Impact: MEDIUM - Affects EMR credibility Decision Needed By: Month 12

Question: Should we pursue HL7 FHIR compliance certification?

Options:

  • A) Yes, pursue certification

    • Cost: $5,000-10,000
    • Benefits: Official certification, trust, partnerships
    • Timeline: 3-6 months
  • B) No, self-certification

    • Cost: $0
    • Benefits: Faster, no cost
    • Limitations: Less credibility

Provisional Answer: Defer to Month 12+, then decide based on:

  • Hospital partnership requirements
  • Market demand for certification
  • Budget availability

Recommended Approach:

  1. Build FHIR R4 support (self-certified)
  2. Test with major EMRs (Epic, Cerner sandboxes)
  3. Month 12: Evaluate need for certification
  4. If needed for partnerships: Pursue certification

Compliance & Security

Q16: GDPR Priority

Priority: P2 (Medium) Impact: MEDIUM - Affects Milestone 5 Decision Needed By: Week 40

Question: Is European deployment a near-term goal? Should we prioritize GDPR compliance?

Options:

  • A) Yes, European deployment planned

    • Implement GDPR features in Milestone 5
    • Right to be forgotten
    • Data portability
    • Consent management
    • Effort: 2-3 weeks
  • B) No, US-only for now

    • Defer GDPR to later (if needed)
    • Focus on HIPAA only
    • Effort: 0

Provisional Answer: Recommend Option B (defer GDPR) unless European deployment confirmed.

Rationale:

  • HIPAA compliance already achieved
  • GDPR can be added later if needed
  • Focus resources on core features first

If GDPR Needed Later:

  • Right to be forgotten: User data deletion API
  • Data portability: Export user data (JSON, CSV)
  • Consent management: Cookie consent, data processing consent
  • Data residency: EU region deployment (optional)
  • Effort: 2-3 weeks

What We Need:

  • Confirmation of European deployment plans
  • Timeline for European launch (if planned)

Q17: Data Residency Options

Priority: P3 (Low) Impact: MEDIUM - Affects architecture Decision Needed By: Month 9

Question: Should we implement data residency options (US, EU, other regions)?

Options:

  • A) Yes, multi-region deployment

    • US region (primary)
    • EU region (GDPR compliance)
    • Other regions as needed
    • Effort: 3-4 weeks per region
  • B) No, US-only

    • Simpler architecture
    • Lower cost
    • Use global CDN for frontend only

Provisional Answer: Recommend Option B (US-only) initially.

Rationale:

  • US-only sufficient for initial launch
  • Global CDN (Vercel/Netlify) provides fast frontend worldwide
  • Backend API can be added to additional regions later if demand

If Multi-Region Needed:

  • Deploy backend to AWS regions (us-east-1, eu-west-1)
  • Use Route 53 for geo-routing
  • Database replication across regions
  • Effort: 3-4 weeks

AI & Machine Learning

Q19: Model Training Strategy

Priority: P2 (Medium) Impact: MEDIUM - Affects accuracy and cost Decision Needed By: Week 20

Question: Should we fine-tune models or rely on prompt engineering only?

Options:

  • A) Fine-tune models

    • Benefits: Higher accuracy, specialized to medical domain
    • Drawbacks: More work, requires training data, GPU infrastructure
    • Effort: 4-6 weeks
  • B) Prompt engineering only

    • Benefits: Faster, no training needed, no GPU required
    • Drawbacks: Less control, lower accuracy
    • Effort: 1-2 weeks
  • C) Hybrid (fine-tune embeddings, prompt engineer LLM)

    • Benefits: Best of both worlds
    • Effort: 3-4 weeks

Provisional Answer: Recommend Option C (hybrid approach).

Recommended Approach:

  1. Fine-tune embeddings:

    • Use BioGPT or PubMedBERT for medical embeddings
    • Improves retrieval quality
    • GPU required (see Q18)
  2. Prompt engineer LLM:

    • Use GPT-4 with optimized prompts
    • No fine-tuning needed
    • Faster, simpler
  3. Benefits:

    • Better retrieval (fine-tuned embeddings)
    • Flexible generation (prompt engineering)
    • Balanced cost and accuracy

Low Priority Decisions (Can Defer)

Design & UX

Q5: Mobile App Strategy

Priority: P3 (Low) Impact: HIGH - Future roadmap Decision Needed By: After Milestone 2 (Month 6)

Question: Should we plan for native mobile apps (iOS/Android) or stick to responsive web/PWA?

Options:

  • A) Responsive web + PWA only

    • Benefits: Single codebase, faster development, lower cost
    • Limitations: Limited native features, slower performance
  • B) Native mobile apps later (React Native or Flutter)

    • Benefits: Better UX, full native features, faster performance
    • Drawbacks: More work, 3 codebases (web, iOS, Android)

Provisional Answer: Recommend Option A (responsive web + PWA) for Year 1.

Rationale:

  • Build excellent responsive web experience first
  • PWA provides app-like experience (install to home screen)
  • Evaluate native apps after Milestone 2 based on:
    • User feedback
    • Mobile usage metrics
    • Feature requests (push notifications, offline mode, etc.)

Decision Point: Month 6 (after Milestone 2)

  • If mobile usage > 40%: Consider native apps
  • If mobile usage < 40%: Continue with PWA

Infrastructure & Operations

Q8: CI/CD Platform

Priority: P3 (Low) Impact: LOW - Most platforms similar Decision Needed By: Week 1

Question: CI/CD platform preference: GitHub Actions, GitLab CI, or CircleCI?

Options:

  • A) GitHub Actions

    • Already using for backend
    • Tight GitHub integration
    • Free for public repos
  • B) GitLab CI

    • If moving to GitLab
  • C) CircleCI

    • Additional cost

Provisional Answer: Recommend Option A (GitHub Actions).

Rationale:

  • Already configured for backend
  • Tight integration with GitHub
  • Free for public repos, generous limits for private
  • Good monorepo support with Turborepo

External Dependencies

Q13: Hospital Partnership Timeline

Priority: P3 (Low) Impact: MEDIUM - Affects EMR integration planning Decision Needed By: Month 6

Question: What's the timeline for hospital partnership discussions?

Provisional Answer: Defer EMR integration to Milestone 7+ (Month 13+).

Recommended Approach:

  1. Focus on core features first (Milestones 1-6)
  2. Month 6: Evaluate market demand for EMR integration
  3. Month 9: Begin hospital partnership discussions
  4. Month 13+: Implement EMR integration if partnerships secured

AI & Machine Learning

Q20: Model Evaluation Framework

Priority: P3 (Low) Impact: MEDIUM - Affects quality assurance Decision Needed By: Week 25

Question: What's the strategy for model evaluation and benchmarking?

Options:

  • A) Manual evaluation by medical experts
  • B) Automated benchmarks (MedQA, PubMedQA, USMLE)
  • C) Hybrid (automated + manual)

Provisional Answer: Recommend Option C (hybrid).

Recommended Approach:

  1. Automated benchmarks:

    • MedQA dataset (US Medical Licensing Exam questions)
    • PubMedQA dataset (biomedical literature questions)
    • Custom test set (100 clinical scenarios)
  2. Manual evaluation:

    • 5-10 medical experts review 100 sample queries
    • Rate accuracy, relevance, completeness
    • Provide feedback on improvements
  3. Continuous evaluation:

    • A/B testing in production
    • User feedback (thumbs up/down)
    • Analytics on query success rate

Q21: Multi-Modal AI Use Cases Priority

Priority: P3 (Low) Impact: MEDIUM - Affects Milestone 6 Decision Needed By: Week 45

Question: Which medical image analysis use cases are highest priority?

Options to Rank:

  • Radiology (X-ray, CT, MRI)
  • Dermatology (skin lesions)
  • Pathology (microscopy)
  • Wound assessment
  • ECG interpretation

Provisional Answer: Recommended Prioritization:

  1. Dermatology (P1)

    • Simpler, well-defined problem
    • Large training datasets available (HAM10000)
    • High accuracy achievable (> 90%)
    • Useful for telehealth
  2. Wound assessment (P1)

    • Useful for telehealth and home care
    • Easier than radiology
    • Good datasets available
  3. ECG interpretation (P2)

    • Requires specialized models
    • Medical liability concerns
    • High value if accurate
  4. Radiology (P3)

    • Most complex, requires specialized training
    • High liability concerns
    • Consider FDA approval path
  5. Pathology (P3)

    • Requires specialized hardware (microscopy)
    • Niche use case

Rationale:

  • Start with simpler use cases (dermatology, wounds)
  • Validate accuracy and user feedback
  • Expand to complex use cases (radiology) later

Decision Template

For each decision, use this template:

### [Question Number]: [Question Title] **Category:** [Design & UX | Infrastructure | External Dependencies | Compliance | AI/ML] **Priority:** [P0 Critical | P1 High | P2 Medium | P3 Low] **Impact:** [HIGH | MEDIUM | LOW] **Decision Needed By:** [Week number or milestone] **Question:** [Clear statement of the question] **Options:** - **A)** [Option description] - Pros: [List] - Cons: [List] - Cost: [Estimate] - Effort: [Time estimate] - **B)** [Option description] - Pros: [List] - Cons: [List] - Cost: [Estimate] - Effort: [Time estimate] **Provisional Answer:** [Recommended option with rationale] **What We Need:** - [ ] [Required information or approval] - [ ] [Additional requirements] **Impact if Not Decided:** [Description of consequences]

Approval Process

Decision Authority

PriorityDecision AuthorityApproval Process
P0 (Critical)Product Owner + CTOWritten approval required
P1 (High)Product OwnerEmail approval
P2 (Medium)Technical LeadTeam consensus
P3 (Low)Development TeamTeam discussion

Approval Workflow

  1. Review Questions: Team reviews all questions
  2. Gather Information: Collect required information
  3. Evaluate Options: Discuss pros/cons of each option
  4. Make Decision: Follow approval process based on priority
  5. Document Decision: Update this document with final decision
  6. Communicate: Notify team of decision and rationale

Progress Tracking

QuestionStatusDecisionDecided ByDate
Q1: Design System✅ ResolvedCreate from scratch (Radix + Tailwind)Product Team2025-11-21
Q2: Storybook✅ ResolvedYes, include in monorepoProduct Team2025-11-21
Q3: Component Library⏳ Pending---
Q4: Dark Mode⏳ Pending---
Q5: Mobile Apps⏳ Pending---
Q6: Deployment✅ ResolvedUbuntu (Docker Compose) initiallyProduct Team2025-11-21
Q7: Staging Env⏳ Pending---
Q8: CI/CD Platform⏳ Pending---
Q9: Telemetry⏳ Pending---
Q10: UpToDate✅ ResolvedNo budget, use free sourcesProduct Team2025-11-21
Q11: API Priorities⏳ Pending---
Q12: EMR Targets⏳ Pending---
Q13: Hospital Partnership⏳ Pending---
Q14: FHIR Cert⏳ Pending---
Q15: Offline PHI✅ ResolvedNo PHI offline, non-PHI onlyProduct Team2025-11-21
Q16: GDPR⏳ Pending---
Q17: Data Residency⏳ Pending---
Q18: GPU Budget✅ ResolvedNo budget, use OpenAI APIsProduct Team2025-11-21
Q19: Model Training⏳ Pending---
Q20: Model Evaluation⏳ Pending---
Q21: Multi-Modal Priority⏳ Pending---
Q22: Image Datasets✅ ResolvedPre-trained models (GPT-4 Vision)Product Team2025-11-21
Q23: AI Liability✅ ResolvedDecision support only, disclaimersProduct Team2025-11-21

Summary: 8/23 questions resolved (all critical questions). Ready for Milestone 1.


Next Steps

  1. Schedule Review Meeting:

    • Team reviews all 23 questions
    • Prioritize critical decisions (8 questions)
    • Assign research tasks for information gathering
  2. Critical Decisions First:

    • Q1: Design system availability
    • Q2: Storybook setup
    • Q6: Deployment strategy
    • Q10: UpToDate licensing
    • Q15: Offline PHI regulations
    • Q18: GPU infrastructure
    • Q22: Medical image datasets
    • Q23: AI diagnosis liability
  3. Gather Required Information:

    • Design assets (logos, colors, fonts)
    • Budget approvals (UpToDate, GPU, telemetry)
    • Legal/compliance reviews (HIPAA, FDA)
    • IT approvals (infrastructure, DNS)
  4. Make Decisions:

    • Follow approval process
    • Document decisions in this file
    • Update roadmap with final decisions
  5. Begin Development:

    • Start Milestone 1, Week 1
    • Revisit open questions at each milestone

Document Version: 1.0 Last Updated: 2025-11-21 Next Review: After critical decisions made (before Week 1) Owner: VoiceAssist Product Team

Beginning of guide
End of guide