Stage 3: EXPAND + MASTER PLAN

# Stage 3: EXPAND + MASTER PLAN
## LUME — Computational Choreography Device

---

3a. RISK AUDIT

### R1: Unity on Jetson Orin Nano Performance [CRITICAL]
- Failure scenario: Unity URP + compute shaders + VFX Graph + video recording exceeds Jetson's 25W GPU budget. Frame rate drops below 30fps, making the experience unusable.
- Probability: MEDIUM-HIGH (40
- Impact: CRITICAL. The entire product is the visual experience. If it stutters, there is no product.
- Mitigation: (a) Build a performance benchmark on actual Jetson hardware in Week 2. Target: 60fps with 256x256 fluid sim, 30K particles, 1080p output. (b) If Unity fails: port the 3 compute shaders to native CUDA/Vulkan. The shaders are standard HLSL and translate directly to CUDA. Lose VFX Graph but gain 3-5x performance. (c) Fallback: drop to 720p output and 128x128 fluid sim. Still looks good on a projector from 6ft.
- Validation criteria: Measured frame time < 16.6ms (60fps) with full pipeline running on Jetson dev kit.
- Owner: Mohamed
- Timeline: Must be resolved by end of Month 1.

### R2: Real-Time Video Compositing + Recording Simultaneously [CRITICAL]
- Failure scenario: Recording composited video (camera feed + visual overlay) to NVMe while rendering the visual pipeline causes frame drops or recording corruption.
- Probability: MEDIUM (30
- Impact: HIGH. Content creation is the primary value proposition. If recording degrades the experience, users choose between "good visuals" and "record content," which breaks the product promise.
- Mitigation: (a) Use Jetson's dedicated hardware video encoder (NVENC) which runs independently of the GPU compute pipeline. H.265 encoding at 1080p30 consumes ~2W and does not steal GPU cycles. (b) Record at 1080p30 even if display is at 1080p60 (30fps is standard for social content). (c) Compositing happens in GPU memory. The final render target is both displayed AND piped to NVENC. No CPU-side copy needed.
- Validation criteria: 5-minute continuous recording at 1080p30 with zero dropped frames while maintaining 60fps display output.

### R3: Femto Bolt Depth Quality at Room-Scale Distances [MEDIUM]
- Failure scenario: At 3-5 meter distance (typical studio), the Femto Bolt's depth accuracy degrades. Silhouette edges become noisy. The fluid sim looks jittery instead of smooth.
- Probability: LOW-MEDIUM (25
- Impact: MEDIUM. Noisy depth can be filtered, but aggressive filtering adds latency.
- Mitigation: (a) Use NFOV mode (640x576 @30fps, 75x65 deg) for tighter, more accurate depth at distance. (b) Apply temporal filtering (exponential moving average across 3-5 frames) to smooth depth edges. (c) The Sobel edge detection in DepthReprojection.compute already provides clean edges from noisy depth. (d) Femto Bolt is rated to 5.5m with <11mm accuracy. At 4m, accuracy is ~15mm. This is acceptable for silhouette-level effects.
- Validation criteria: Clean silhouette extraction at 4m distance with no visible jitter on 1080p display.

### R4: Echelon Audio Engine Port to aarch64 [MEDIUM]
- Failure scenario: The 17-crate Rust Echelon engine does not compile or does not run correctly on aarch64 Linux (Jetson's architecture). Dependencies may assume x86.
- Probability: LOW-MEDIUM (20
- Impact: MEDIUM. GENERATE mode is the copyright-free moat. Without it, LUME is in LISTEN-only mode, which still works but loses the strongest differentiator.
- Mitigation: (a) Rust cross-compilation to aarch64-unknown-linux-gnu is well-supported. Test early. (b) Audio synthesis is pure math (oscillators, filters, envelopes). No x86-specific code. (c) Fallback: pre-generate audio tracks on cloud (Suno API) and download to device. GENERATE mode becomes "download royalty-free tracks matched to your movement" rather than real-time synthesis.
- Validation criteria: Echelon engine running on Jetson dev kit producing audio output at 44.1kHz with <10ms latency.

### R5: BOM Cost Exceeds $700 Target [MEDIUM]
- Failure scenario: Femto Bolt ($418) + Jetson ($249) + NVMe ($40) + cameras ($20x2) + mic array ($15) + speakers ($10x2) + PCB + enclosure + cables + thermal = $850-950. Margin disappears at $1,299 retail.
- Probability: MEDIUM (35
- Impact: MEDIUM. Either raise retail price to $1,499 (acceptable) or cut components.
- Mitigation: (a) Negotiate Orbbec volume pricing. At 500+ units, Femto Bolt should be $350-380. (b) Use Femto Mega ($399 MSRP) if available at lower volume pricing (same iToF sensor, different form factor). (c) Drop speakers from V1 (save $20) and require external audio. (d) Use 1080p cameras instead of 4K (save $15 per camera). 1080p is sufficient for social media content.
- Validation criteria: BOM < $750 at 500-unit volume. BOM < $650 at 3,000-unit volume.

### R6: FCC/CE Certification Delays [MEDIUM]
- Failure scenario: The device contains active radio (WiFi 6E, Bluetooth) and must pass FCC Part 15 and CE/RED testing. Testing reveals EMI issues from compute board + cameras + speakers in close proximity. Requires PCB redesign.
- Probability: MEDIUM (30
- Impact: MEDIUM-HIGH. Cannot legally sell in US/EU without certification. 3-6 month delay.
- Mitigation: (a) Use Jetson Orin Nano Super as a module. NVIDIA's module is pre-certified (FCC/CE/IC). Only the carrier board + enclosure need testing. (b) Femto Bolt is pre-certified as a peripheral. (c) Engage a certification consultant (e.g., Element/Intertek) from Month 2, not Month 6. (d) Budget $30K for testing + consultant.
- Validation criteria: Pre-scan pass at Month 4. Formal certification by Month 7.

### R7: Content Sharing Pipeline Latency [LOW-MEDIUM]
- Failure scenario: After a session ends, the AI editor takes 5+ minutes to produce the 60-second reel. The dancer leaves the studio before content is ready. The share-to-social moment is lost.
- Probability: MEDIUM (30
- Impact: MEDIUM. Every minute of delay reduces the probability of sharing by ~20
- Mitigation: (a) The AI editor runs DURING the session, not after. It continuously identifies "best moments" based on energy spikes, pose triggers, and audio beats. When the session ends, it only needs to stitch the pre-identified segments. (b) Target: content ready in < 60 seconds after session end. (c) WiFi Direct transfer to phone: 300 Mbps for 100MB file = 2.7 seconds.
- Validation criteria: Reel available on companion app < 90 seconds after session end.

### R8: Multi-Person Experience Quality [LOW-MEDIUM]
- Failure scenario: With 3+ people in the depth field, the fluid sim becomes chaotic. Too many motion vectors, too much density injection. The visual output is a noisy mess instead of beautiful art.
- Probability: MEDIUM (30
- Impact: LOW-MEDIUM. V1 can target solo/duo use. Multi-person is a V1.5 feature.
- Mitigation: (a) Per-person depth segmentation using distance thresholding (people at different distances get separate fluid fields). (b) Total force injection capped regardless of person count (divide force by number of people detected). (c) Multi-person mode in PresetManager: different palettes per person, interaction zones where their fluid fields merge.
- Validation criteria: Visually coherent output with 3 people at 3m distance.

### R9: One-Person Company Scaling Bottleneck [MEDIUM]
- Failure scenario: Hardware support requests (defective units, setup issues, firmware bugs) consume all of Mohamed's time. Development stops. No new features, no marketplace, no growth.
- Probability: HIGH (50
- Impact: HIGH. Without continued development, the product stagnates.
- Mitigation: (a) Self-service support: comprehensive video setup guides, in-app diagnostics, OTA firmware updates. (b) AI-powered support: LLM-based chatbot trained on LUME knowledge base handles 80
- Validation criteria: <10

### R10: Competitor Fast-Follow [LOW]
- Failure scenario: After LUME launches and demonstrates market fit, a well-funded startup or big tech (NVIDIA, Apple, Meta) launches a competing product within 12 months.
- Probability: LOW (15
- Impact: MEDIUM. First-mover advantage in content + marketplace + community provides 12-18 month moat.
- Mitigation: (a) Content network effects: by the time a competitor launches, LUME has 50K+ pieces of content, 1000+ visual packs, and an established community. (b) Keep innovating: V2 hardware, new visual capabilities, AI choreography, social platform. (c) Patents: file provisional on the composite content pipeline (visual overlay + royalty-free audio + auto-edit). (d) The one-person cost structure means LUME can be profitable at a scale where competitors are still burning cash.
- Validation criteria: 18-month head start on content/community before first credible competitor appears.

---

3b. EXPANDED SPECIFICATIONS

Spec 1: Hardware Bill of Materials

Retail: $1,299 (500 qty = 34% margin) / $1,299 (3000 qty = 45
Or: $1,499 at launch (500 qty = 43

Spec 2: Software Architecture (Jetson)

┌─────────────────────────────────────────────────────────┐
│                     LUME OS (L4T Ubuntu)                  │
│                                                           │
│  ┌──────────────────────────────────────────────────┐    │
│  │              Unity Runtime (URP, IL2CPP)           │    │
│  │                                                    │    │
│  │  DepthSource ──► DepthProcessor ──► OpticalFlow   │    │
│  │       │                                    │       │    │
│  │       │              FluidSimulator  ◄─────┘       │    │
│  │       │                    │                       │    │
│  │  AudioAnalyzer ──► AudioToFluid                    │    │
│  │       │                    │                       │    │
│  │       ▼                    ▼                       │    │
│  │  ColorPaletteManager   VFX Graph (particles)      │    │
│  │       │                    │                       │    │
│  │       └──────────┬─────────┘                       │    │
│  │                  │                                 │    │
│  │           Render Pipeline (URP)                    │    │
│  │                  │                                 │    │
│  │    ┌─────────────┼─────────────┐                   │    │
│  │    │             │             │                   │    │
│  │  HDMI Out    Compositor    Frame Recorder          │    │
│  │  (display)   (overlay on   (NVENC H.265)          │    │
│  │              camera feed)                          │    │
│  └──────────────────────────────────────────────────┘    │
│                                                           │
│  ┌──────────────────┐  ┌────────────────────────┐        │
│  │  Echelon Engine   │  │  Content Manager        │        │
│  │  (Rust, aarch64)  │  │  (Rust)                 │        │
│  │                    │  │                          │        │
│  │  Audio synthesis   │  │  AI Director (Murch)    │        │
│  │  5 genre kits      │  │  BreathPhase pacing     │        │
│  │  Motion-reactive   │  │  Moment detection       │        │
│  │  Royalty-free      │  │  Auto-edit (reel/clip)  │        │
│  │                    │  │  WiFi AP for phone share │        │
│  │  44.1kHz output    │  │  NVMe session management │        │
│  │  to speakers +     │  │  OTA firmware updates    │        │
│  │  HDMI + 3.5mm     │  │                          │        │
│  └──────────────────┘  └────────────────────────┘        │
│                                                           │
│  ┌──────────────────────────────────────────────────┐    │
│  │  Device Manager (Rust systemd service)             │    │
│  │  - Boot sequence (attract mode auto-start)        │    │
│  │  - OTA update client                              │    │
│  │  - WiFi AP mode for phone pairing                 │    │
│  │  - Health monitoring + telemetry                   │    │
│  │  - Depth camera lifecycle                         │    │
│  │  - NVMe space management (auto-cleanup old sessions)│   │
│  └──────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────┘

Spec 3: Companion App (iOS/Android)

LUME Companion App

SCREENS:
1. Device Pairing (WiFi Direct / QR code)
2. Live View (real-time preview from LUME, ~200ms delay)
3. Control Panel:
   - Visual preset selector (12+ options, visual thumbnails)
   - Color palette picker
   - Audio mode toggle (LISTEN / GENERATE)
   - Genre/BPM selector (for GENERATE mode)
   - Recording controls (start/stop/screenshot)
4. Content Gallery:
   - Recent sessions (thumbnails with metadata)
   - AI-edited reels and highlights
   - Raw footage access
   - Export/share to TikTok, IG, YouTube, Camera Roll
5. Audio Studio:
   - Suno AI track generator (genre, mood, BPM, duration)
   - Track library (previously generated + marketplace)
   - Download tracks to LUME for offline use
6. Marketplace:
   - Visual packs (browse, preview, purchase)
   - Audio packs
   - Scene templates
   - Artist profiles
7. Settings:
   - Device name, WiFi, firmware version
   - Display output (resolution, aspect ratio)
   - Camera settings (exposure, white balance for content cams)
   - Storage management
   - Account / subscription

TECH STACK:
- SwiftUI (iOS) / Kotlin Compose (Android)
- WebSocket to LUME for real-time control
- WiFi Direct for content transfer
- RevenueCat for subscriptions
- Stripe for marketplace transactions
- Supabase for user accounts + content metadata

Spec 4: Content Compositor Pipeline

The compositor is the critical path for LUME's value proposition. It runs on the Jetson GPU and produces the final content files.

INPUT STREAMS (simultaneous):
  1. Wide cam: 1920x1080 @ 30fps (H.264 from camera ISP)
  2. Tight cam: 1080x1920 @ 30fps (portrait, H.264)
  3. Femto RGB: 1920x1080 @ 30fps (center, from depth camera)
  4. Rendered frame: 1920x1080 @ 60fps (visual engine output)
  5. Audio: 44.1kHz stereo (from mic array or Echelon)

COMPOSITOR OPERATIONS:
  1. For each camera feed, every frame:
     a. Decode camera H.264 frame (NVDEC hardware decoder)
     b. Alpha-blend rendered visual frame over camera frame
        - Visual frame has black background where no particles
        - Additive blend mode: camera + particles = final pixel
        - Blend factor varies by visual preset (some presets
          have dense particles that need transparency adjustment)
     c. Encode composited frame to H.265 (NVENC hardware encoder)
     d. Write to NVMe session buffer

  2. Parallel: AI Director continuously scores each camera angle
     - Maintains "best moments" index (timestamp + score + camera)
     - Detects pose triggers (ArmsUp, ArmsSpread, etc.)
     - Tracks audio beats/transients for cut timing

  3. On session end:
     a. Read "best moments" index
     b. Select top 60 seconds for reel (portrait cam, 9:16)
     c. Select top 30 seconds for highlight (multi-cam, 1:1)
     d. Apply cuts on beats (from audio transient timestamps)
     e. Encode final reel and highlight (NVENC)
     f. Generate thumbnail from highest-energy frame
     g. Signal companion app: content ready for transfer

PERFORMANCE TARGET:
  - Compositing: <2ms per frame (GPU shader operation)
  - NVENC encoding: <1ms per frame at 1080p30
  - Session end to content ready: <60 seconds
  - WiFi transfer: <10 seconds for 100MB file

Spec 5: Attract Mode + Installation UX

LUME's first impression determines adoption. The attract mode must be compelling enough that a person walking into a studio says "what IS that?" and then immediately understands how to interact.

ATTRACT MODE STATE MACHINE:

  [BOOT] ──► [CALIBRATING] ──► [ATTRACT] ──► [ACTIVE] ──► [RECORDING]
                                    │              │            │
                                    │              ▼            ▼
                                    │         [IDLE] ──►  [ATTRACT]
                                    │              │
                                    │              ▼
                                    └──────── [SLEEP] (display off, 30 min idle)

ATTRACT:
  - Display: breathing particles, slow color cycle (12s per palette)
  - Audio (GENERATE mode): ambient pads, very quiet (10% volume)
  - LED ring: slow pulse in current palette primary color
  - Depth camera: scanning for body entry into field

TRANSITION TO ACTIVE:
  - Trigger: depth silhouette detected (body enters 0.5-4m zone)
  - Duration: 500ms crossfade
  - Visual: particles accelerate toward detected silhouette
  - Audio: rhythm layer fades in
  - LED ring: brightens to full

ACTIVE:
  - Full visual pipeline running
  - Content cameras recording (always, circular buffer of last 10 min)
  - Audio reactive (LISTEN) or generative (GENERATE)
  - Pose triggers enabled

TRANSITION TO IDLE:
  - Trigger: no motion for 2 seconds (IdleDetector.cs logic)
  - Visual: fluid freezes (SuperHot mode)
  - Audio: drums mute, pad-only
  - After 5 seconds: transition to ATTRACT

TRANSITION TO RECORDING:
  - Trigger: user starts explicit recording (app button, voice, or gesture)
  - Visual: subtle "REC" indicator in corner of display
  - LED ring: solid red
  - All cameras recording dedicated files (not just circular buffer)
  - Session metadata logging begins

---

3c. MASTER EXECUTION CHECKLIST

### Wave 0: PREFLIGHT (Month 1, Weeks 1-2)
Validate all critical assumptions before committing to hardware development.

Wave 0 Kill Criteria: If task 0.5 fails (cannot achieve 60fps on Jetson), evaluate: (a) reduce to 30fps target, (b) port to native Vulkan, (c) switch to Jetson AGX Orin ($999, 275 TOPS). If (c) is required, BOM increases by $750 and retail must be $1,999+. Decision point.

---

### Wave 1: CORE ENGINE (Month 1-2, Weeks 2-6)
Build the visual engine running on Jetson hardware.

Wave 1 Gate: Full visual pipeline running on Jetson at 60fps with depth camera input. A person can stand in front of LUME and see reactive particles on a connected display.

---

### Wave 2: CONTENT ENGINE (Month 2-3, Weeks 6-10)
Build the recording, compositing, and auto-editing pipeline.

Wave 2 Gate: Complete content loop. Dance in front of LUME. End session. Get a 60-second reel with visual overlay and audio, ready to share, within 90 seconds.

---

### Wave 3: AUDIO ENGINE + APP (Month 3-4, Weeks 10-14)
Integrate Echelon, build companion app, make the experience end-to-end.

Wave 3 Gate: End-to-end experience. Dance with LUME. Get content. Share to social. Audio is royalty-free. App controls everything. OTA updates work.

---

### Wave 4: HARDWARE DESIGN + PROTOTYPING (Month 4-6, Weeks 14-22)
Design the physical product and build hand-assembled prototypes.

Wave 4 Gate: 3 prototypes deployed in real studios. Beta feedback collected. CM engaged. No showstopper issues.

---

### Wave 5: PRODUCTION + LAUNCH (Month 6-9, Weeks 22-36)
Manufacturing, certification, and market launch.

Wave 5 Gate: Product in market. Customers using LUME. Revenue flowing. Support manageable.

---

### Wave 6: MARKETPLACE + GROWTH (Month 9-12, Weeks 36-48)
Build the platform layers that drive recurring revenue and network effects.

Wave 6 Gate: Marketplace live with 20+ packs. Subscriptions generating recurring revenue. Second production run shipped. Monthly revenue > $100K.

---

MONTH-BY-MONTH ROADMAP (YEAR 1)

MONTH 1: PROVE IT WORKS
├── Purchase Jetson + Femto Bolt (Week 1)
├── Port Unity project to Jetson (Week 1-2)
├── Performance benchmark (Week 2) ← KILL GATE
├── Cross-compile Echelon (Week 2-3)
├── Trademark + domain (Week 1)
└── Optimized engine running at 60fps (Week 3-4)

MONTH 2: MAKE IT RECORD
├── NVENC recording module (Week 5-6)
├── Content compositor (Week 6-7)
├── AI Director port (Week 7-8)
├── Auto-editor (reel + highlight) (Week 8)
└── Demo: dance → 60s reel in 90 seconds

MONTH 3: MAKE IT COMPLETE
├── Echelon GENERATE mode integration (Week 9-10)
├── Companion app v1 (pairing + control + gallery) (Week 10-12)
├── Suno API integration (Week 11)
├── OTA update system (Week 12)
└── Demo: full end-to-end, phone app to TikTok post

MONTH 4: MAKE IT REAL
├── Industrial design engagement (Week 13-14)
├── Carrier board design (Week 14-16)
├── 3D-printed prototype enclosures (Week 15-16)
├── 3 hand-assembled prototypes (Week 16)
└── Deliverable: 3 working LUME prototypes

MONTH 5: MAKE IT TESTED
├── Deploy 3 prototypes to beta studios (Week 17)
├── 2-week beta period (Week 17-18)
├── Feedback analysis + priority fixes (Week 19)
├── Pre-certification EMI scan (Week 19-20)
├── CM engagement + quotation (Week 20)
└── Decision: production run sizing + pricing

MONTH 6: MAKE IT CERTIFIED
├── FCC/CE certification testing (Week 21-24)
├── Finalize production PCB (Week 22)
├── Enclosure tooling ordered (Week 23)
├── Website build (lume.dance) (Week 22-24)
├── App Store submission (Week 24)
└── Pre-orders open on website

MONTH 7: MAKE IT MANUFACTURED
├── First articles from CM (Week 25-26)
├── QC validation (Week 27)
├── Production run begins (500 units) (Week 27-28)
├── Packaging design + production (Week 26-28)
├── Launch content creation (10 demo videos) (Week 26-28)
└── Production units arriving at fulfillment center

MONTH 8: MAKE IT LAUNCHED
├── Ship first 100 units (Week 29-30)
├── Launch PR campaign (Week 29)
├── Social media push (demo videos) (Week 29-32)
├── First customer feedback collection (Week 30-32)
├── Hotfix firmware updates as needed (Week 30-32)
└── Target: 100 units shipped, <5% support rate

MONTH 9: MAKE IT GROW
├── Ship remaining 400 units (Week 33-36)
├── Marketplace development begins (Week 33)
├── Subscription tiers activated (Week 34)
├── First 5 visual packs from external artists (Week 35)
├── Events rental pilot launch (Week 36)
└── Target: 500 units sold, $649K hardware revenue

MONTH 10: MAKE IT A PLATFORM
├── Marketplace launch (20 visual packs) (Week 37-38)
├── API/SDK documentation (Week 38-40)
├── Analytics dashboard (Week 39-40)
├── Second production run ordered (2,000 units) (Week 40)
└── Target: $19/month subscribers reaching 200

MONTH 11: MAKE IT SCALE
├── Multi-device sync feature (Week 41-42)
├── Events rental expansion (3 new markets) (Week 42-44)
├── Developer beta program (10 developers) (Week 43-44)
├── LUME Social prototype (Week 43-44)
└── Target: 300 subscribers, $80K MRR (hardware + subs)

MONTH 12: MAKE IT DEFENSIBLE
├── Second production run ships (2,000 units) (Week 45-48)
├── LUME Social beta launch (Week 46)
├── Year 1 retrospective + Year 2 planning (Week 48)
├── Patent filing (composite content pipeline) (Week 47)
└── TARGET: 700+ units sold, 400+ subscribers, $50K+ MRR recurring

---

PULSE AUTO-SPAWN ASSESSMENT

Tasks tagged `automate: true` in the checklist above that could be dispatched as autonomous agent sessions:

**26 of 48 tasks (54

---

KEY INSIGHT

LUME is not a hardware product. LUME is a content creation platform that happens to require hardware to exist. The hardware is the distribution mechanism for a software platform that generates unique, royalty-free, visually stunning content from physical movement.

The Peloton parallel is exact: the bike was never the product. The community, the content, the instructors, the leaderboard were the product. LUME's equivalent: the visual effects, the audio engine, the marketplace, the social feed, the content sharing loop.

The hardware gets LUME into the room. The content keeps it there. The social platform makes it spread.

"Move in light."