Skip to content

Working prototype · Baku ID 2026 · Offline-first Edge AI

SeetheWorld.HearWhatMatters.

AI-powered smart glasses that help blind people navigate safely, understand surroundings, read text, recognize familiar faces, and make independent decisions in real time.

100% of safety works offline556 automated tests<300ms danger interruptRaspberry Pi 5

The Problem

Sight loss shouldn't mean losing independence

White canes feel one metre ahead. Guide dogs cost $50,000+. Premium devices like OrCam cost $4,500 and still don't warn about an open manhole.

285M

people worldwide live with visual impairment — 39M fully blind (WHO).

Every day

Open manholes, stairs and traffic are daily life-threatening hazards in cities.

The gap

Existing aids describe objects — they don't make navigation decisions.

Mosttechnologydescribestheworld.Ablindpersoncrossingastreetneedsmorethandescriptiontheyneedadecision.VVAwatches,remembers,weighsthedanger,andsaystheonesentencethatmatters.Calmly.Intime.Evenwithnointernetatall.

The Solution

A co-pilot for the streets, not a narrator

VisionVoiceAssist fuses computer vision, a cognitive decision layer and natural Azerbaijani speech into one wearable.

Sees & prioritises

80+ object classes with stable tracking. In a 20-object scene it speaks max 2 calm sentences — danger first, always.

Life-saving hazard layer

Stairs, holes, open manholes (circular-signature detection), crosswalks, curbs — fully on-device, 3-frame confirmation, zero-false-alarm tested.

Decides, not describes

“Obstacle ahead — pass on your left.” Clock directions, step distances, wait/stop commands. A fail-safe engine blocks risky advice.

Reads the world

Signs, menus, price tags in AZ/EN/RU/TR. Stops-and-reads automatically; prices spoken in words.

Recognises loved ones

“Your mother is in front of you.” Encrypted on-device face database — biometrics never leave the glasses.

Interrupts for danger

A critical warning cuts any speech in under 300 ms — measured by automated tests, with pre-cached emergency audio.

The Hardware

Inspect the prototype. Every angle.

Drag to rotate · scroll to zoom · tap the glowing points to open each module. This is the actual product geometry — not an illustration.

Built on Raspberry Pi 5

Quad-core edge compute inside the temple — no phone, no tether.

Edge AI

Detection, hazards, OCR, faces and speech run fully on-device.

Offline Safety

Zero cloud dependency for anything that keeps the user safe.

Live AI — not a mockup

Talk to the brain. Feed it your world.

This is the same Gemini vision brain the prototype runs when online. Upload a photo and hear exactly what the glasses would say — or just have a conversation with VVA.

LIVE
Frame received
Detecting objects
Reading text
Judging hazards
Composing guidance
The glasses' verdict appears here —
objects, hazards, text, and the exact sentence they would speak.

Photos are resized in your browser, analysed once, never stored. Free-tier AI — if it rate-limits, give it a minute.

Real-World Scenarios

Press play on a day in the life

Each scenario replays the exact module chain the glasses execute — same order, same rules, same words.

Interactive Architecture

Click any stage. See why it exists.

Nine isolated, tested modules on one event bus — from photons to a calm Azerbaijani sentence.

Camera

USB/CSI capture with automatic reopen — 3 consecutive timeouts trigger announcement + recovery loop, never a crash. File/RTSP sources supported for testing.

System Telemetry

An operating system for the street

Replay of desktop-session telemetry (values animated for display — measured baselines: 13.4 FPS camera, 10.7 ms CV layer, 88 MB RAM, <300 ms danger interrupt).

VVA RUNTIME · TELEMETRY REPLAY OFFLINE-READY LIVE LOOP

Camera FPS

13.4

CV layer / frame

10.7ms

AI confidence

91%

Objects tracked

6

Hazards

0

Speech queue

1

RAM

88MB

Danger latency

240ms

Offline-First

Works anywhere on Earth. No internet required.

The entire safety layer — detection, hazards, traffic lights, OCR, face recognition, speech — runs on-device. Cloud AI (Gemini → Claude fallback chain) only adds richer scene descriptions when available. Lose connection mid-street? The glasses announce it and keep protecting you.

  • Models pre-downloaded at install time
  • Automatic online/offline switching in <2s
  • Local TTS chain: Piper → pyttsx3 → espeak → beep
  • Encrypted local face & place memory

100%

safety features offline

<2s

offline failover

4

OCR languages

0

cloud dependency for danger

Technology

Battle-tested stack, edge-optimised

YOLOv8n
ByteTrack
OpenCV
Gemini
Claude
Whisper
Piper
Tesseract OCR
Python
Raspberry Pi 5
Edge AI
React
Next.js
Three.js
0

automated tests — all passing

0ms

max danger-interrupt latency

0+

object classes recognised

0

critical CVEs (pip-audit)

Privacy & Safety

Trust is the core feature

Faces stay on-device

Biometric embeddings are Fernet-encrypted locally. Frames sent to cloud are face- and plate-blurred first — provable via the live dashboard.

Threat-modelled

10 attack vectors documented with mitigations and test evidence. Secrets masked in logs. Dashboard is localhost + token only.

Fail-safe by design

After any hazard, movement advice is blocked for 6 seconds — the system says “stop”, never a risky direction. Uncertain? It stays silent.

Market Opportunity

A $7B assistive-tech market with no affordable leader

285M visually impaired people globally · ~70,000 in Azerbaijan alone · B2C device sales + B2G social-support programs. The i18n-centric core makes Turkish, Russian and Georgian markets a module away.

VisionVoiceAssistOrCam MyEyeWhite Cane
Price~$200 BOM$4,500$40
Hazard warnings (manhole, stairs)✓ on-device1m reach only
Navigation decisions✓ “pass left / stop”✗ describes only
Works offline✓ 100% of safetypartial
Azerbaijani (native-quality)✓ first-class
Face recognition (encrypted, local)✓ (cloud-linked)
Learns from user feedback

Engineering Quality

Due diligence welcome. The code holds up.

Not a hackathon script — a release candidate with CI-grade evidence behind every claim on this page.

0

automated tests, all passing

0%

line coverage

0

lint findings (ruff, 414 fixed)

0

known CVEs (pip-audit)

Documented

Architecture map, developer guide, threat model, security setup, release notes with honest limitations, phase reports.

Hardened

Threat-modelled (10 vectors + mitigations), secrets masked in logs, encrypted biometrics, localhost+token dashboard.

Operable

systemd watchdog, OTA updates with rollback, endurance & FPS benchmark tooling, doctor command for field diagnostics.

Meet the Founder

Built end-to-end by one determined founder

Aliasgar Fatullayev testing the VVA prototype: wearing the camera glasses connected to a Raspberry Pi 5 while the vision pipeline reads a juice box on the laptop screenPROTOTYPE · LIVE TEST

Aliasgar Fatullayev

Founder & Engineer · 14 y.o. · Baku, Azerbaijan

Designed and shipped the entire system — vision pipeline, cognitive layer, speech chain, hardware integration, security model and this website — through 10 engineering phases with test evidence at every gate.

LinkedIn
AI / Computer VisionPythonEdge AIEmbedded SystemsCybersecurityProduct DesignSystem ArchitectureTypeScript / React
  1. Started programming self-taught in Baku

  2. First Raspberry Pi experiments

  3. Computer vision & edge AI deep-dive

  4. VVA architecture: offline-first, fail-safe by design

  5. Working prototype — 556 automated tests

  6. Desktop edition running live

  7. Baku ID 2026 — stage demo

  8. Mission: independence for every blind person

Roadmap

From prototype to platform

  1. Working prototypenow

    556 passing tests · desktop edition live · v1.0 release candidate tagged

  2. Baku ID 2026now

    Live stage demo: hazards, traffic light, OCR, face recognition, offline failover

  3. Pilot with blind community

    20-user field trial in Baku with the Society of the Blind; feedback-learning loop

  4. Hardware rev. 2

    Custom frame, ToF distance sensor, bone-conduction audio, 8h battery

  5. Regional expansion

    Turkish & Russian language packs — the core is already i18n-first

For Investors

Real device. Real code. Real users waiting.

Seeking seed funding for hardware rev. 2, the 20-user Baku pilot and certification. Full technical due-diligence pack available: architecture, threat model, test reports.

Working prototype

Runs today on Raspberry Pi 5 and any Windows laptop — same core.

Offline edge AI

No cloud bills per user. No connectivity risk on stage or street.

Pilot-ready

Systemd watchdog, OTA updates with rollback, 60-min endurance tooling.

Founder velocity

Built by a 14-year-old founder from Baku — 10 engineering phases shipped with test evidence.

Future Pilot Partners

Who we're building the pilot with

Conversations planned for the 20-user Baku field trial. Shown as categories — no signed partnerships are claimed yet.

Society of the Blind

IN DISCUSSION · PLACEHOLDER

Special Schools

IN DISCUSSION · PLACEHOLDER

Hospitals & Clinics

IN DISCUSSION · PLACEHOLDER

Universities

IN DISCUSSION · PLACEHOLDER

NGOs & Foundations

IN DISCUSSION · PLACEHOLDER

City Accessibility Programs

IN DISCUSSION · PLACEHOLDER

Pitch Deck

The full story, investor-ready

INVESTOR VERSION · 2026

VisionVoiceAssist

See the World. Hear What Matters.

Working prototype · Baku ID 2026

  • Problem → solution → live prototype evidence
  • Architecture, threat model & test reports included
  • Market sizing: 285M worldwide · ~70k in Azerbaijan
  • Roadmap to 20-user pilot and hardware rev. 2
31 slides · 1.1 MB Updated 2026 Due-diligence pack on request

FAQ

Judges ask. We answer.

Why Raspberry Pi 5 instead of a phone?

A phone occupies the hand a blind user needs for their cane. Pi 5 integrates into the glasses, drives a haptic motor over GPIO, and runs the full vision stack locally for a ~$200 bill of materials — 20× cheaper than OrCam.

What happens when internet drops?

Nothing critical. Detection, hazards, traffic lights, OCR and speech are all on-device. The glasses announce “switched to offline mode” within 2 seconds and keep going; cloud descriptions resume automatically.

How do you prevent false alarms?

Floor hazards need 3 consecutive confirming frames; identical alerts are rate-limited; a 100-frame noisy-corridor test must produce zero alarms in CI. Below 35% confidence the system deliberately stays silent.

Is the face data safe?

Only 512-dimension mathematical embeddings are stored — never photos — encrypted with Fernet on the device. Frames leaving for cloud description are blurred first. One command deletes a person forever.

Is this a demo or a product?

It's a release candidate: 556 automated tests, zero lint findings, zero known CVEs, systemd watchdog, OTA with rollback. Experimental parts (behaviour prediction, voice queries) are explicitly labelled as such.

Contact

Let's give sight a voice — together

founder@visionvoiceassist.com github.com/Orest002/VVA LinkedIn — Aliasgar Fatullayev

Demo requests & investor conversations welcome. Based in Baku, Azerbaijan.