Mayday — A voice-activated lifeline when seconds matter.

1 · One-liner

Mayday is a voice-triggered personal-safety app. A user speaks a short phrase, the phone silently flips into an evidence-recording and SOS-broadcast mode, and an AI voice agent on a 24/7 contact-center line engages with the threat environment on the user's behalf — even if the phone is snatched, locked, or smashed.

2 · Problem

The moment of danger

When someone is threatened (assault, carjacking, domestic-violence escalation, kidnapping, medical emergency alone), they have 2–8 seconds to act. Fumbling for the phone, unlocking it, opening an app, and pressing a button is exactly the workflow the threat makes impossible. Existing personal-safety apps require a tap; hardware panic buttons require a hand that may be pinned; voice assistants are passive.

Why current solutions fail

• Tap-required apps (bSafe, Noonlight, Citizen) — useless if hands are pinned or phone is grabbed.
• Hardware buttons (Revolar, InvisaWear) — require a deliberate, visible gesture; one device only; no escalation if snatched.
• Passive voice assistants (Siri "call 911") — generic, no pre-staged evidence, no broadcast, no contact-center, no chain of custody.
• Bluetooth wearables — single point of failure, battery, and discoverable.

What the user actually needs in those 8 seconds

• Trigger without hands, without looking at the phone.
• Silent — no obvious indication the phone just flipped into emergency mode.
• Irreversible evidence capture (front + back + mic) uploaded off-device instantly.
• Multi-channel broadcast to contacts & social.
• A human on the other end of a voice call who can hear the scene, talk to the attacker, and dispatch.

3 · Users & segments (US launch)

4 · Capabilities

4.1 Core feature set (MVP, V1)

4.2 V2 (post-MVP)

• Smartwatch extension — Wear OS / watchOS app for tap OR voice; survives phone theft.
• Mesh relay — nearby Mayday users become silent BLE relay if cell is jammed.
• AI scene classification — on-device audio classifier detects gunshots, screams, breaking glass to auto-trigger.
• Insurance discount integration — Lemonade, MetLife affinity discounts (proven conversion lift on telematics apps).
• Family / group plans — parents monitoring teens, B2B2C enterprise.

5 · Experience flows

5.1 Onboarding (≈90 seconds)

1. Permission ask sequence (each prompt has a 10-second animated explainer + "Why we need this" panel before the OS dialog): (1) microphone — drives the wake-word listener; (2) location — Always Allow — drives the broadcast map + 911 dispatch; (3) camera — dual-cam evidence; (4) microphone + media-recording — Android MediaProjection / iOS mic-indicator shown to user during capture; (5) motion & fitness — fall/impact supplement; (6) notifications + critical alerts (iOS) — broadcast delivery. Foreground-only location is rejected with a one-tap "Open Settings" CTA; the app remains unarmed until Always is granted.
2. User records trigger phrase × 3 (anti-spoof) and safe-word.
3. User adds 3–10 trusted contacts (with phonebook friction bypass).
4. User pre-authors 1–3 social blast templates.
5. Run a 30-second drill mode that simulates the event end-to-end (does not dial 911).

5.2 Trigger → dispatch (happy path, ~20s end-to-end)

5.3 Operator console (web dashboard)

• Live audio + video stream (WebRTC), GPS, sensor timeline.
• User profile side panel: medical notes, emergency contacts, photo.
• One-click 911 dispatch with location + audio relay (via RapidSOS / Prepared / Carbyne integration).
• Transcription overlay + AI-suggested next questions ("Are you injured?", "Are you alone?").
• Multi-event queue, supervisor escalation, post-event evidence export.

6 · Architecture

6.1 Stack

7 · AI voice & contact-center model

The PRD's most distinctive bet: when the phone may already be stolen, the user is not the voice caller. Mayday itself places an outbound call from a Twilio number to the user's phone (and to contacts) and an AI voice agent speaks with the attacker on the user's behalf.

7.1 Three-tier call flow

1. Tier 1 — AI voice agent (seconds 0–30): greets, identifies as "Mayday safety assistant on a recorded line for [user name]", continues to listen, transcripts, and escalates based on threat signals (raised voices, weapon mention, location data).
2. Tier 2 — Human supervisor handoff (< 30s target): trained de-escalation agent takes over via LiveKit handoff. AI summarizes context, continues to log & record.
3. Tier 3 — 911 dispatch: human agent uses RapidSOS / Carbyne to push GPS + audio relay to the nearest PSAP. Many PSAPs accept silent relay today (Text-to-911 is universal; audio relay is in 60% of CA, expanding).

7.2 Build vs. buy

7.3 Unit economics (per event)

Events assumed at 0.6 / user / month (1 every ~2 months per active user); see §10.

8 · Market

8.1 Demand signals (US, 2024–2026)

• Crime anxiety: 71% of US women say they feel unsafe walking alone at night (Pew, 2024); up from 57% in 2019.
• Rideshare: ~3.8M US gig drivers; 1 in 4 report a safety incident in the last 12 months (ITUC survey).
• Campus: 1 in 5 women experience sexual assault in college; $400M/yr spent on campus safety tech.
• DV: 16M US adults experience intimate-partner violence annually (CDC); < 5% use a safety app.
• AI voice cost collapse: Twilio voice minutes dropped 35% YoY in 2024–25; LLM voice ~6× cheaper than 2023.
• Smartphone baseline: 92% US smartphone penetration; iOS 17 / Android 14 ship on > 80% of in-use devices.

9 · Competitive landscape

9.1 Defensibility / moat

• 24/7 contact-center ops — capital-intensive, slow to replicate.
• Audio-event ML (gunshot / scream classifier) trained on labeled safety data.
• Brand trust — first-mover in voice-activated safety; PR moat after a high-profile saved life.
• B2B2C distribution (universities, rideshare platforms, insurers).
• Network data — every incident improves threat-detection models; data flywheel.

10 · Business model & path to $1M ARR in 12 months

10.1 Revenue model

10.2 12-month plan to $1M ARR

10.3 Sensitivity & downside

11 · Risks & mitigations

12 · Legal & compliance

• Recording consent: all-party consent (CA, FL, IL, MD, MA, NH, PA, WA, MT). Default to opt-in disclosure ("This call may be recorded for safety") during AI voice.
• COPPA: under-13 not eligible; 13–17 with parental consent flow.
• HIPAA: if integrating health data, Business Associate Agreement + minimum-necessary principle.
• Two-party consent (CA, FL, etc.): AI voice agent must announce recording at handoff to human.
• Terms of Service: Mayday is a safety enhancement, not a replacement for 911; user acknowledges no guarantee.
• Export control: encryption export review (encryption is exempt from EAR for normal use but reviewed).
• Insurance: professional liability + cyber + media liability, ~$25k/yr at this scale.
• Law-enforcement data requests: process within 72h, log all requests, public transparency report.

13 · Success metrics (north-star + ladder)

14 · Roadmap (12 months)

15 · Open questions / decisions needed

1. BPO choice — US-based boutique (Citizen's "Protect" partner model) vs. international for cost? Trade-off: voice quality & trust vs. COGS.
2. Brand name — keep "Mayday" (good recall, but FAA-trademark-adjacent) or rename? Candidates: Sentinel / Resound / Vesper / Clearview.
3. Always-on vs. session-mode default — battery vs. protection. Ship both; default to session + 8-hr scheduled auto-arm.
4. Funding — bootstrap to first $250k ARR (no contact-center scale) or raise $2–4M seed for 24/7 ops + AI training data?
5. Distribution partner priority — Uber / Lyft (B2B) vs. university safety (B2B2C) vs. insurance (B2C at scale).

Recommended next step: validate C7 (AI voice call) + BPO with a 2-week Lo-fi spike (Twilio + OpenAI Realtime + a 5-seat BPO), and confirm with 20 rideshare-driver interviews that the AI-voice-to-stranger framing works in a real threat scenario. Estimated: $4k, 2 weeks.

Appendix · What this PRD demonstrates

• Capability: clear scope, feature table, architecture diagram, end-to-end flow with timing budget.
• Viability: market sizing with sources, $1M ARR plan quarter-by-quarter, unit economics per event, CAC/LTV, sensitivity.
• Honesty: risks section names the things that will kill this — false triggers, two-party consent, iOS background restrictions, PSAP audio-rejection, liability.
• Differentiation: the AI-voice-from-our-number + human contact-center model is the one piece no one else ships; the rest of the market is tap-triggered.