Product Requirements Document¶
HumanityAI'd Museum AR Smart Glass Guide — v1.0 · July 2026
Confidential — HumanityAI'd
Download
1. Product overview¶
The Museum AR Smart Glass Guide turns a museum visit into a hands-free, personalized, multilingual tour. Visitors wear lightweight AR glasses (RayNeo X3 Pro); the glasses stream what the visitor is looking at to an on-premises AI server, which recognizes the exhibit in front of them and instantly plays a professionally narrated audio guide in the visitor's language, while contextual text and imagery can be shown on the glasses' display.
Everything — visual recognition (CLIP + FAISS), translation content, and neural text-to-speech (Chatterbox Multilingual) — runs inside the museum's own network. No visitor images or behavioural data leave the building: a decisive advantage for government and cultural-sector procurement in the region.
2. Problem statement¶
| Today's pain | Consequence |
|---|---|
| Handheld audio guides require numbered keypads and constant map-checking | Visitors look at devices, not art |
| Guided group tours run on fixed schedules and 1–2 languages | Lost revenue from non-Arabic/English speakers |
| Mobile apps demand downloads, drain phones, and depend on visitor hardware | < 15% typical adoption |
| Analytics on visitor flow and engagement are coarse or absent | Curators cannot measure which exhibits work |
3. Target users & personas¶
- The Visitor — international tourist; wants zero-friction, own-language narration that starts automatically when they look at an exhibit.
- The Curator / Content Manager — owns exhibit content; needs a web panel to add exhibits, photos, translations, and to review/approve AI-generated narration before it goes live.
- The Operations Manager — issues and collects glasses at the front desk; needs device/session management and daily analytics.
- The Museum Director / Digital Transformation lead — buys on visitor experience, data sovereignty, and measurable engagement.
4. Core user journeys¶
- Wear & go: visitor receives glasses, staff starts a session with language preference → visitor walks; when they look at an exhibit for ~2 s, recognition fires and narration plays in their language. No buttons.
- Curated tours: predefined routes (e.g., "Highlights in 60 minutes") with stop-by-stop guidance and tour progress tracking.
- Content lifecycle: curator uploads 1–3 photos per exhibit → system builds visual index automatically → writes (or auto-generates) descriptions → generates TTS narration per language → reviews and approves each clip → live.
- Analytics review: dwell time per exhibit, most/least visited pieces, audio play-through rates, per-session tour summaries.
5. Functional requirements¶
5.1 Visitor experience (glasses app)¶
- F-1 Hands-free exhibit recognition from live camera stream (WebRTC/H.264; target < 1.5 s from gaze to result)
- F-2 Automatic audio narration playback in session language; 8 launch languages: en, ar, fr, es, zh, ru, hi, de (Arabic incl. RTL text display)
- F-3 On-lens display of exhibit title and key facts; adjustable via web-based glasses settings panel
- F-4 Offline-tolerant behaviour: reconnect logic when Wi-Fi drops; session resumes
5.2 Recognition service¶
- F-5 CLIP ViT-L/14 image embeddings + FAISS similarity index; multiple photos per exhibit (angles) for robustness
- F-6 Recognition threshold configurable per deployment; returns exhibit + translation + approved audio URL in one response
- F-7 Index updates live on image upload — no restart or re-training required
5.3 Content management (admin panel + API)¶
- F-8 CRUD for museums → galleries → exhibits (categories, inventory numbers, artist, period, tags)
- F-9 Multi-image upload with primary/angle designation; automatic embedding
- F-10 Translation records per exhibit per language: title, short & full description, fun fact, audio script
- F-11 One-click TTS generation per exhibit across all languages (female/male voices); approval workflow — narration is served to visitors only after explicit review approval
- F-12 Tours: ordered stops, multilingual tour metadata
5.4 Sessions & analytics¶
- F-13 Session lifecycle (start/end, language, device), event logging (view, dwell, audio played)
- F-14 Analytics: overview dashboard, popular exhibits, per-session visit summary with dwell seconds
- F-15 Prometheus metrics + Grafana dashboards for ops (GPU, latency, request rates)
5.5 Platform & deployment¶
- F-16 Fully on-premises Docker deployment (single GPU server, NVIDIA runtime); all services containerized (API, PostgreSQL, Redis, nginx, Prometheus, Grafana)
- F-17 API-key security for all endpoints; admin login for panel; per-key rate limits
- F-18 Data persistence across restarts for images, FAISS index, audio, and DB (named volumes)
6. Non-functional requirements¶
| Requirement | Target |
|---|---|
| Recognition latency (frame → result) | < 1.5 s p95 |
| Recognition precision on enrolled exhibits | > 95% with 1–3 photos/exhibit |
| Narration quality | Neural TTS, museum-tuned loudness (−18 LUFS), Opus 80 kbps |
| Concurrent visitors per GPU server (8 GB) | 25–40 active sessions |
| Availability | 99.5% during opening hours; watchdog + auto-restart |
| Privacy | No cloud egress of visitor imagery; sessions pseudonymous |
| Languages | 8 at launch; TTS engine supports 23 (expansion-ready) |
7. Proven reference deployment (July 2026)¶
An end-to-end pilot runs today at the HumanityAI'd office ("HumanityAI'd Office" museum instance): 10 exhibits, 12 enrolled photos, 8 languages, 80 approved narrations, generated and served by the exact stack above on a single RTX 4060 Ti (8 GB). Enrolment-to-live time for a new exhibit: under 5 minutes including narration generation.
8. Out of scope (v1)¶
- Visitor-owned-device (BYOD phone) mode — roadmap v1.5
- Indoor positioning / wayfinding overlays — roadmap v2
- Face recognition of visitors (deliberately excluded — privacy posture)
9. Success metrics¶
- ≥ 60% of equipped visitors complete ≥ 5 exhibit narrations
- Average dwell time on narrated exhibits +30% vs. baseline
- Content team can enrol a new exhibit in < 10 minutes without vendor help
- NPS of glasses experience ≥ 50
10. Dependencies & risks¶
| Risk | Mitigation |
|---|---|
| GPU memory contention (recognition + TTS on one 8 GB card) | TTS generation is an offline/admin operation; runtime serving is pre-generated audio. 16 GB GPU recommended for large catalogues |
| Glasses hardware supply (RayNeo) | App isolates vendor SDK; portable to other Android-based AR glasses |
| Long Chinese scripts increase TTS memory | Script-length guideline (≤ 60 s narration) enforced in authoring UI |
| Museum Wi-Fi quality | Site survey in POC phase; dedicated SSID recommended |