# Interview: PetCare (Smart Multi-Pet Scheduler) **Role:** Intermediate Developer (3–4 yr Android → Web) **Mockup:** `petcare-mockup.html` **Duration:** ~60 min --- ## Setup (5 min) Show mockup. Say: > "This is a smart multi-pet care scheduler. A household with multiple pets manages feeding, medication, walks, and health tracking in one dashboard. There's an AI advisor for health questions and a pet sitter mode. You need to build this as a web application. Let's talk through the design." --- ## Part 1: Sub-system Decomposition (15 min) > **"How would you break PetCare into sub-systems or services? Walk me through the architecture."** ### What they should surface: | Sub-system | Key Responsibility | |------------|-------------------| | **Scheduling Engine** | Recurrence rule evaluation, task instance generation, overdue detection, snooze logic. Core complexity — harder than meal planner | | **Pet Profile Service** | CRUD per pet, species-specific fields, medical history, weight tracking | | **Care Task Service** | Task definitions (template with recurrence), per-pet dosing, supply tracking | | **Health Records Service** | Vet visits, vaccinations, weight history, photo attachments | | **Notification Service** | Push reminders at task time, overdue escalation (yellow → red), sitter notifications | | **Sitter Access Service** | Token generation, expiry, access scope enforcement, audit log | | **AI Advisor Service** | Symptom triage, care suggestion, health scan analysis. Thin orchestration layer over LLM API | | **Inventory Service** | Food, medication, supply tracking — auto-refill detection | ### Push deeper: > **"The scheduling engine needs to handle 'insulin every 12 hours' AND 'heartworm pill first Monday of every month' AND 'walk twice a day but skip if raining.' How do you model recurrence rules?"** Look for: RRULE / iCalendar format awareness, exception handling, timezone sensitivity (insulin time doesn't shift with DST). Strong candidate will say "use a library" but also describe data model and querying: pre-generate instances for near future or compute on-the-fly with a cron-like evaluator. > **"Where do overdue calculations live — are they computed on read or persisted as state?"** Good answers: persisted as state (status field on TaskInstance), updated by a lightweight tick. Read-time computation works for small scale but breaks with 50+/day tasks per household. Offline editing complicates read-time evaluation. --- ## Part 2: Identify System Challenges (12 min) > **"Pick two sub-systems. What's the hardest technical challenge in each — the one that could fail in production?"** ### Key challenges to surface: **Challenge: Recurrence engine at scale** - Pre-generate task instances for N days ahead, or compute on-the-fly and cache? - If pre-generate: how far ahead? 7 days? 30 days? What happens when user changes recurrence — retroactive or forward-only? Re-generate all future instances or diff? - Exception rules compound: "Every day except vet days" → recurrence + exception set - If on-the-fly: query for a date range involves iterating rules × pets — slow at 1000+ households **Challenge: Overdue detection with offline users** - User marks "done" offline at 8am. Another user's device shows it overdue until sync completes. - Conflict: user A marks done offline at 8am, user B marks overdue and re-snoozes at 8:05. Sync at 8:10 — whose state wins? - Good answer: last-write-wins with activity log for audit. Or: task goes to "completed" only when both agree (problematic for sitter handoff) **Challenge: Multi-species care normalization** - Dog feeding: 2 cups kibble, twice daily. Cat: ½ can wet food + dry available all day. Rabbit: unlimited hay + daily veg + water change. - How does one data model represent these? Species-specific templates with a common "care action" abstraction layer? **Challenge: Sitter access security** - Share link → anyone with URL can access household schedule. URL shared in WhatsApp → leaks. PIN adds friction. - JWT in URL parameter = logged in server logs, referer headers. Good: hash fragment (#) or magic link flow (one-time token, server-side validation + immediate invalidation on use). - Expiry must be server-enforced, not client-only. **Challenge: AI health scanner liability** - Photo shows a lump. AI says "likely benign." User delays vet visit. Turns out malignant. - Who is liable? How to disclaim without undermining trust? - Good answer: clear UI disclaimers, hard fallback thresholds (AI never says "diagnosis"), language matching urgency to recommend vet visit. Vet referral gate at certain confidence levels. ### Push for opinion: > **"If you have limited engineering resources, which challenge do you solve first — the recurrence engine or the sitter access security?"** Look for: trade-off reasoning, not just "both." Good candidate weighs: security defect = trust/reputation damage, recurrence defect = poor UX. Recurrence is harder to retrofit. Security is easier to bolt on but harder to fix after a breach. --- ## Part 3: Coding Agent / AI Best Practices (10 min) (Same structure as MealPlan — re-use for consistency. Ask again if they gave different answers in MealPlan interview. Note: if this is a separate interview with different candidate, use standalone.) > **"Let's talk about using AI coding tools to build PetCare. What's your approach?"** ### Watch for: **Patterns that carry over from MealPlan OR show new depth:** - AI for generating test fixtures for the recurrence engine (date math edge cases = great AI task) - AI for writing the RRULE parser wrapper (well-defined input/output = great AI task) - AI for generating the token-validation middleware for sitter access (security-critical → human review mandatory) - AI for generating UI components across 3 pet types (good — repetitive patterns) - AI for nutrition/medication dose calculations (bad — high-risk numeric logic, hallucinations matter) **PetCare-specific AI best practices:** - AI-written sitter token generation → MUST reseed RNG properly. Common AI mistake: `Math.random()` for security tokens - AI-written overdue logic → date math across timezone boundaries. Timezone-naive code is very common AI mistake - AI for CRUD boilerplate across 8 entities → good. High leverage, low risk. - AI for health record schema → review carefully. Labelling fields incorrectly (e.g. "last_vaccination" vs "next_due") is common. ### If they haven't used coding agents: > **"Your teammate generated the sitter access middleware with Claude Code. As reviewer, what do you inspect?"** Look for: token randomness quality, expiry enforcement (server-side, not trusting client claims), rate limiting, audit logging, revoked token check. --- ## Part 4: Building an AI Feature (15 min) > **"Design the Health Scanner AI feature. User takes a photo of their pet's skin/eye/ear issue. System helps assess urgency. Walk me through UX, model, data flow, error handling."** ### UX Layer: - **Entry:** button "Scan Health Issue" in health records or AI sidebar. Camera opens in-app (not system camera, to add guidance overlay). - **Guidance overlay:** frame the affected area, good lighting icon, hold steady. Close-up guide. - **Processing states:** 1. "Analyzing photo..." (0–3s) 2. "Checking for known patterns..." (3–6s) 3. Results appear incrementally: detected features first, then assessment, then care steps - **Confidence UX:** - High confidence item → displayed directly with green checkmark - Medium → yellow badge, "likely" language - Low → "I'm not sure — here's what it could be" with list - **Result card:** photo thumbnail + detected condition name + description + urgency level (routine/watch/vet-visit/emergency) + care steps + "When to see a vet" section - **Actions:** Save to health records, set follow-up reminder, share with vet (download/email summary) - **Fallbacks:** - Dark/blurry photo → "Please retake with better lighting" - Partial obstruction (fur blocking) → "I can only see part of the area — try parting the fur" - No match → "I can't identify this. When in doubt, consult your vet." + option for human remote vet consultation (v2) - Network failure → "Can't process photo right now. Save draft and try again later." ### System Architecture: ``` ┌─────────────────────────────────────────────┐ │ [Camera] → Image Compressor (max 4MB) │ │ ↓ │ │ Moderation Check (NSFW / gore filter) │ │ ↓ │ │ Multimodal LLM (Claude Vision / GPT-4o) │ │ Prompt: "Describe the skin condition shown │ │ in this pet photo. Be specific about │ │ appearance, location, severity indicators. │ │ DO NOT give a diagnosis." │ │ ↓ │ │ Structured Output Extraction │ │ {features[], suggested_conditions[], │ │ urgency, confidence, care_steps[]} │ │ ↓ │ │ Post-processing: │ │ - Urgency classifier override (safety net) │ │ - Disclaimers injected │ │ - Vet threshold check → referral prompt │ │ ↓ │ │ UI rendering │ └─────────────────────────────────────────────┘ ``` ### Key considerations they should surface: 1. **Model choice:** Multimodal LLM over fine-tuned CNN. More flexible, handles wide variation (species × body part × lighting × breed). Cost per scan is the trade-off. 2. **Image moderation:** Pet photos shouldn't need NSFW filter, but zoomed-in skin issues can look medical/graphic. A filter prevents edge cases from reaching the model. 3. **Structured output:** Don't let LLM free-text the diagnosis. Use schema-enforced output (tool use / function calling) to get consistent fields for UI rendering and safety checks. 4. **Safety net layer:** AI says "likely allergy." But image matches "ringworm (fungal)" with 30% confidence — urgency classifier should still flag it "recommend vet" even if primary suggestion is low-urgency. 5. **Fallback escalation path:** System fails at photo analysis → offer text symptom input instead. "Tell me what you see — redness? swelling? discharge? how long?" 6. **Compliance:** Veterinary telemedicine regulations vary by jurisdiction. Features that recommend treatment may be regulated. Must consult legal before rolling out. 7. **Feedback loop:** "Was this helpful?" after each scan → used to improve prompts, not retrain model. Track: accepted vs dismissed suggestions, follow-up visits. 8. **Bias across species:** Model trained mainly on dog/cat photos → rabbit/guinea pig skin issues may perform worse. Need per-species confidence calibration. ### Edge case test: > **"User takes photo of their cat. AI says 'likely allergic dermatitis.' They skip the vet. Two weeks later it's worse. Who is at fault?"** Strong: "The app needs to communicate this is not a diagnosis — it's a triage tool. Language matters: 'consistent with allergic dermatitis. See your vet for diagnosis.' Hard fallback thresholds ensure ambiguous or high-urgency cases default to vet recommendation. The photo result screen must not look like a medical report — it should look like a preliminary assessment with prominent referral call-to-action." --- ## Part 5: UI/UX Critique (if time, 5 min) > **"Look at the mockup again. What would you change?"** ### What to look for: - **Multi-pet filter clarity:** Pet tabs filter schedule but leave health panel showing Cooper's data when Luna is selected — this is inconsistent - **Notification overload:** 3 pets × 8 tasks = 24 notifications/day. Need quiet hours, per-pet notification toggle, or batch summary - **Timeline density:** When all 3 pets in one timeline, rows are dense. Some tasks are per-pet, some shared. How to visually distinguish? - **Task grouping:** "Breakfast — Cooper & Luna" is grouped in mockup but "Luna's insulin" is separate. Is the grouping deterministic or user-configurable? - **Missing sitter view:** No mockup of what sitter sees — important to show simplified interface - **Overdue vs upcoming visual distinction:** Overdue tasks (red) and upcoming (blue) are clearly different, but what about late-night tasks? Dark mode consideration? - **Empty state:** First-time user adds pet → empty timeline? Could show sample schedule to demonstrate value - **Accessibility:** Timeline uses color-only urgency (red/yellow/green). Colorblind users need patterns or text labels. --- ## Scoring Rubric | Area | Minimum (Hire Signal) | Strong (Accelerate Signal) | |------|----------------------|---------------------------| | Decomposition | 4+ sub-systems with clear boundaries, mentions scheduling complexity | 7+, explains data flow between services, describes pre-generate vs compute-on-fly tradeoff for recurrence | | Challenges | 2+ challenges with plausible approach | Identifies recurrence engine pre-generation vs on-fly, sitter token expiry enforcement, multi-species normalization | | Coding Agent | Has used AI tools, reviews output, describes specific practice | Articulates "danger zones" (security tokens, timezone math, numeric logic), tests AI-generated code differently by domain risk | | AI Feature | End-to-end flow with UX states, model choice, disclaimers | Mentions moderation gate, structured output extraction, safety net layer, species bias, compliance, feedback loop | | UI/UX Sense | 2+ improvements with rationale | Spots inconsistency (pet filter doesn't change health panel), notification overload, sitter view absence, accessibility gaps | --- ## Decision Flow ``` Can they design? → Part 1 & 2 → ≥4 sub-systems + 2 challenges with depth Can they build? → Part 4 → full flow with safety and fallback layers Can they use AI? → Part 3 → specific practices, risk-aware Can they design UX? → All parts → do they surface multi-pet complexity without prompting? Hire if at least 3/4 "minimum" with 1 "strong." Strong hire if 3+/4 "strong." ```