On-Device AI for Museum Experiences

MOSTRAmi turns any phone into a private, multilingual museum guide that explains the artwork in front of you - without sending photos to the cloud.

CHANNELS: iOS / Android / WeChat

•

COMPLIANCE: EU AI Act / GDPR

↓ Press Down for TRL status & Web Simulator

01-D

TRL Status & Live Demo

MOSTRAmi is currently an Android TRL 5-6 internal demo. The v0.8.0 (21) AI Integrity source candidate was compiled on August 14: Android tests and the R8 YC build pass, as do Functions 8/8, token Worker 7/7 and mostrapack 9/9. The artifact is unsigned; the current signed v0.7.0 private-beta APK remains the only distributable build for authorised testers and is also retained as the rollback artifact for future v0.8 qualification. Claim-level citations, multilingual Clio Live, image-first detail exploration, Scholar and museum experience packs are represented across app and site.

TRL 5-6 (Relevant Environment): Android source and the current signed v0.7.0 private-beta APK demonstrate scan context, source-grounded chat, local/cloud routing, detail crops and guided experience packs.
Live Audio Preview: Clio Live targets Firebase Gemini Live native audio and exposes runtime retry inside the Live panel.
Remaining Gate: the 0.8.0 artifact still needs signing, cloud canary, physical-device qualification, Play-track distribution and user evidence.

📱

Interactive Web Simulator

Experience the complete on-device client, offline VLM steering, and BLE radar right in your browser.

Open Simulator ➔

02

The Problem

Museum visitors get a bad guide experience, while museums face high build barriers.

Highlights-Only: Traditional guides cost EUR 5-10 per visit, are language-limited, and cover only 40% of the collection - leaving most artwork with no commentary.
Mandarin Gap: 175M Chinese outbound tourists/yr ($280B spend) [S1] find patches in Mandarin guide coverage and cannot access Western app stores in Europe.
CMS Constraints: Existing guide apps depend on museum-authored CMS. If an artwork isn't in their database, you get nothing.
High B2B Barriers: Building a digital guide costs museums EUR 30-80K and takes 6-12 months of manual content production.

↓ Press Down for B2B Cost & Server OPEX Analysis

Tuscany 2026

Language & Coverage Gaps

Labels and audio guides fail to cover the long tail of museum collections.

Static printed cards only

Standard guides

Clunky Hardware BYOD

Physical guide rentals or cloud apps that lose connection behind thick concrete walls.

EUR 5-10 per-visit rentals

02-D

Cloud LLM Cost Bottleneck

Cloud API Cost Breakdown (Linear Scaling)

Every photo analysis and audio guide generation sent to cloud VLMs scales costs linearly with visitor traffic.

Cost per Scan (GPT-4o/Gemini Pro)~EUR 0.05

Daily Scans per Active User4 scans

10,000 Visitors/dayEUR 2,000 / day

Monthly Cloud OPEXEUR 60,000 / mo

MOSTRAmi On-Device Edge Solution

Local inference runs on the visitor's hardware accelerator. The cost curve is flat, and cloud dependency is eliminated.

Cost per Scan (Local Gemma)EUR 0.00

In-App Model Storage Size~1.6 GB

10,000 Visitors/dayEUR 0.00 / day

Monthly Cloud OPEXEUR 0.00 / mo

03

The Solution

MOSTRAmi runs on a hybrid stack: Gemma 4 LiteRT for private local fallback, Gemini Cloud for deep interpretation, and Gemini Live native audio for voice-first conversation. Point phone, scan artwork, lower the device, then speak with Clio while looking at the work.

Zero Cloud Reliance: No photos leave the device. Complete GDPR compliance by default, bypassing concrete walls blocking internet.
Scan-Any-Artwork: AI generates context on-the-fly, explaining uncatalogued objects that Bloomberg/Smartify CMS apps miss.
Western Channels: Native iOS + Android apps targeted at Western cultural tourists.
Chinese Channel: WeChat Mini Program + Huawei AppGallery (HMS adapters) to bypass GMS blocks without HarmonyOS NEXT rewrites.

↓ Press Down for LiteRT-LM Architecture & Code

iOS AppKMM

Camera Scan Active

iOS Premium

Android AppLiteRT

TTS Offline synthesis

Android Mass

WeChat MPMandarin

No PlayStore install

Chinese Outbound

Real-Time B2B SaaS Dashboard for Curation & Analytics

03-D

LiteRT-LM & GPU Inference

Native Mobile GPU/NPU acceleration using Google's LiteRT-LM runtime.

Local Inference: Under the hood, `LiteRTLM` pre-loads the compiled Gemma 2B model directly onto the smartphone NPU.
Response Latency: Token streaming begins in < 350ms on mid-range devices (Snapdragon 7/8, Google Tensor G3/G4).
Fallback Logic: An internal AI Router detects hardware performance. If NPU accelerations are missing, it handles seamless fallback to low-cost cloud routes.

// Initialize LiteRT Edge LLM val options = LlmInferenceOptions.builder() .setModelPath("/data/local/tmp/gemma-2b-it.bin") .setMaxTokens(512) .setTemperature(0.2f) .build() val inference = LlmInference.createFromOptions(context, options) val response = inference.generateResponse(prompt)

04

Why Now

Three independent enablers just converged to make MOSTRAmi possible in 2026, after being technically and commercially unfeasible before.

Gemma 4 + LiteRT-LM 0.16: local inference is now practical on selected supported devices, but latency, memory, thermal and multilingual quality remain device-matrix gates.
EU AI Act in Force: Cloud-based photo-sending apps face strict regulatory hurdles (GDPR facial capture of bystanders). On-device processing complies by design.
China Outbound Rebound: Outbound tourism surged back to 175M trips and $280B/yr [S1], creating massive demand for guides on GMS-free phones.

↓ Press Down for WebSocket TourSync™ Code

2023

Cloud-only VLMs. High latency, unviable server COGS, cellular dropouts.

2025

Initial on-device ML benchmarks. VLM accuracy low, language support limited.

2026

Gemma 4 LiteRT GA + EU AI Act Enforcement + China reopens outbound.

04-D

WebSocket TourSync™ Engine

Clio TourSync™ syncs guides and visitors in real time over secure WebSocket.

Steering Broadcast: The guide steers the VLM narration (e.g. changing style to "Kids" or "Academic") and sends the artwork ID.
Local Audio Playback: Each visitor phone receives the payload, triggers the local VLM, and streams synthetic voice synchronously.
Low Latency: WebSocket frames deliver in < 50ms across standard B2B museum network interfaces.

// WebSocket steering connection session.incoming.consumeEach { frame -> if (frame is Frame.Text) { val payload = Json.decodeFromString<SyncPayload>(frame.readText()) if (payload.action == "STEER_PERSONA") { viewModel.changeNarrativeStyle(payload.persona) viewModel.triggerLocalTTS(payload.artworkId) } } }

05

Market Size: Bottom-Up

We refuse to cite refuted market aggregator reports. We build our addressable ceiling using bottom-up ICOM/EGMUS counts and SmartGuide-anchored benchmarks.

Dense Art Market: Italy alone records 254M international tourist nights/yr [F1].
Target B2B Museums: ICOM lists ~5,000 Italian museums. Our 3-year addressable target focuses on the top 1,000 museums in the 50K-500K visitors/year band.
Price Band: Anchored to F2 SmartGuide published SaaS tiers of EUR 100 / 190 / 390 / 590 per month.
Italy B2B Ceiling: 1,000 museums × EUR 390 midpoint × 12 = EUR ~4.7M ARR addressable in Italy alone, before expanding B2C.

↓ Press Down for WeChat Outbound Strategy Details

5,000 ICOM Italian Museums

1,000 Mid-Size Target (50k-500k visitors/yr)

Validation target: one exhibition PoC + one tested paid offer; no ARR assumed

05-D

WeChat Mini-Program Strategy

No Western App Stores

Outbound Chinese tourists cannot access Google Play Store in Europe. Installing custom APKs faces device warnings.

WeChat Integration

Our VLM and OCR functions are wrapped in a WeChat Mini-Program. Tourists scan a physical QR code at the museum entrance and open the guide instantly.

Huawei AppGallery

Outbound distribution of the Android native client is also published directly on Huawei AppGallery, bypassing Google Mobile Services.

06

Market Drivers: PESTEL Analysis

P Political & Regulatory EU-wide push for cultural heritage digitization (PNRR) backed by €500M in Italian allocations [F4]. Government focus on regional museum access.

E Economic COGS Edge Zero-server COGS on-device model routing. Bypasses the uncapped OPEX of cloud API queries, making software scale-up affordable for public museum budgets [F2].

S Social & Inbound Shift Chinese outbound tourism surge (~175M trips) [S1] with high intent to visit EU museums. Severe language barrier and lack of GMS playstore access [S4].

T Technological Enablers Gemma 4 GA + mobile NPU hardware parity in 2026. Local vision-language inference and Piper TTS run offline behind thick museum concrete walls.

E Environmental / Green On-device local processing reduces data center cooling and network overhead. Fits green computing mandates for EU public-sector procurement.

L Legal GDPR Exemption EU AI Act (Art. 50) compliant. Local face-blurring ensures no raw biometric bystander pictures leave the client's phone, eliminating museum data liabilities.

↓ Press Down for GDPR Face Blur Code Pipeline

06-D

GDPR Face Obfuscation

Camera frames undergo automatic face blur on the device before any VLM processing.

Local Detection: Google ML Kit Face Detection runs on the camera frame buffer locally.
Instant Blur: Detected bounding boxes are filled with a Gaussian blur matrix in **< 15ms**.
Act 50 Compliance: Ensures no biometrics are exported or analyzed in the cloud, fulfilling EU AI Act compliance.

// On-device canvas blur pipeline val faces = faceDetector.process(image).await() faces.forEach { face -> val box = face.boundingBox canvas.drawRect(box, Paint().apply { shader = BlurMaskFilter(15f, Blur.NORMAL) }) }

07

Business Model

1. B2B SaaS Tiers

EUR 100-590/mo

2. B2C Pro Sub

EUR 4.99/mo

3. B2B Per-Visitor

EUR 0.30-1.00

4. White-Label

€15k + €1.5k/mo

5. Ticketing Comm.

5-8% Fee

Revenue Stream Breakdown

B2B SaaS subscription (primary): 4 tiers anchored to SmartGuide (EUR 100 / 190 / 390 / 590 / month). No Enterprise 5th tier above EUR 590 [F2].
B2C subscription (secondary): Pro at EUR 4.99/mo or EUR 29.99/yr, Family at EUR 9.99/mo for up to 5 devices. iOS carries a 91% ARPU premium [S5].
B2B per-visitor (alternative): EUR 0.30-1.00 per visitor per destination, up to 70% volume discount [F2].
White-label (opportunistic): Setup fee + monthly licensing [illustrative].
Ticketing referral: 5-8% commission via Tiqets/Musement [industry estimate].

↓ Press Down for Detailed 3-Year Financial Model Table

07-D

Evidence Gates

Gate	Window	Required evidence	Commercial claim
Current state	August 2026	v0.8.0 source candidate; current signed v0.7.0 private-beta APK; 15 private testers	0 paying customers; EUR 0 revenue
Closed beta	Sep 2026-Jan 2027	Activation, interruption, multilingual voice, stability, retention	No public launch or revenue target
Exhibition PoC	Feb-Apr 2027	One completed, scoped deployment if agreement is finalized	PoC validation; not counted as revenue in advance
Release decision	May-Sep 2027	Open-beta/public-release decision and one tested paid offer	No ARR assumed

08

The Competition

Why MOSTRAmi wins: We bypass the structural limitations of CMS content, cloud liabilities, and app-store access blocks.

1. CMS-Free Autonomy (Generative vs. Pre-Authored)

Bloomberg Connects and legacy wands require curators to write scripts for every single object. MOSTRAmi's on-device VLM explains any artwork dynamically - even if the museum has zero catalog entries.

2. GDPR-by-Design Privacy (Offline vs. Cloud Upload)

Smartify and Google Arts upload images to external servers, capturing bystander faces and risking GDPR/AI Act compliance fines. MOSTRAmi uses local face blurring and runs fully offline behind stone walls.

3. Chinese GMS-Free Wedge (WeChat vs. App Store Blocks)

Chinese tourists carry Google Services-free phones, blocking Play Store downloads. We bypass this entirely with a Mandarin WeChat Mini Program and AppGallery APK, tapping into the €280B cohort [S1].

↓ Press Down for Feature-by-Feature Comparison Matrix

CMS-Bound Scan-Anything (VLM) On-Device Private Cloud Dependent

MOSTRAmi (🏆)

Bloomberg

Smartify

Google Arts

SmartGuide

08-D

Competitive Comparison Grid

Feature	Smartify	SmartGuide	Bloomberg Connects	MOSTRAmi
VLM Image recognition	Cloud-Only	Cloud-Only	✗ No	On-Device (Offline)
Real-Time Audio Chat	✗ No	✗ No	✗ No	✓ Clio Live (Voice)
Custom B2B SDK	Only Web-view	Static audio	Walled Garden App	✓ WebSocket TourSync™
China Outbound Ready	✗ No	✗ No	✗ No	✓ WeChat Mini-Program

09

The Team

Fabrizio Degni

Solo Founder & Architect

Built the current Android prototype solo.
Experienced mobile and ML engineer, with background covering on-device inference and localization systems.
Deep first-hand context on Italian art networks and regional museum operations.

Location: Italy (Relocating to SF for batch in-person)

↓ Press Down for 18-Month Execution Roadmap

Team Reality Check

Fabrizio is currently a solo founder. Most top accelerators prefer co-founding teams of 2 or more.

Top use of pre-seed funding will be recruiting and hiring a technical co-founder within the first 60 days of the accelerator batch.

09-D

Milestones & Roadmap

Sep-Oct 2026: Closed Beta

Begin invite-only Android access only after signing, cloud canary and exact-device gates are green. Run at least 30 real-context sessions and measure first scan, second scan, multilingual voice, interruption, latency, crashes and willingness to pay; this is not a public launch.

Nov 2026-Jan 2027: Extended Beta

Expand the cohort and device matrix gradually, harden support and pack handling, and produce a measured beta learning report.

Feb-Sep 2027: PoC and Release Decision

Complete one scoped exhibition PoC, test the first commercial offer, then decide whether evidence supports an open beta or limited public Android release.

Raising USD 500,000

Standard Deal: $125k on 7% post-money SAFE + $375k uncapped MFN SAFE

1. Co-Founder Hire

Technical co-founder sourced and hired during first 60 days of the SF batch.

2. Senior Engineer

Build stabilization and CI first. iOS/KMM remains demand-gated research until Android usage and delivery capacity justify the port.

3. EU Pilot Acquisition

Sales travel, partner scoping and localization to complete the first exhibition PoC and test a first paid B2B offer by September 2027.

12-MONTH COMMERCIAL GATE: one completed PoC + one tested paid offer by September 2027; no ARR is assumed before agreement and delivery.