ALOK

An AI assistant for visually impaired and blind people — built for Bangladesh.

📺 Watch the demo · Top 100 Teams — SOLVIO AI Hackathon 2025 (Team pip_install_chaos)

---

Why

Over 650,000 adults in Bangladesh are blind, and most assistive technology is built for English-speaking users. People who are visually impaired struggle to recognize objects around them, identify Bangladeshi currency notes, read printed text without Braille, and avoid collisions with nearby obstacles.

ALOK ("আলোক" — light) addresses this with real-time object detection and spoken feedback in Bangla, designed to run on standard consumer hardware.

What it does

• Dual-model object detection — every frame is run through both a pretrained YOLOv11 model (80 general object classes) and a custom YOLO model trained on our own Bangladeshi currency dataset, so the system recognizes everyday objects and Taka notes simultaneously.
• Proximity awareness — detections are filtered by bounding-box area, so only objects close enough to matter are announced, with a configurable proximity zone and cooldown between announcements.
• Bangla voice announcements — detected objects are mapped to Bangla names through a built-in translation dictionary and spoken aloud using Google TTS (bn-BD), e.g. "সামনে মানুষ এবং গাড়ি".
• REST API + live mode — a FastAPI backend exposes detection over HTTP for integration, alongside a live webcam mode with on-screen annotations.

Demo

▶️ Product demo on YouTube

The live view shows both models running at once — blue boxes for the custom currency model, green for the pretrained model — with real-time Bangla audio announcements.

How it works

backend/
├── main.py                  # FastAPI app entry point
├── requirements.txt
├── detection/
│   ├── model.py             # YOLOModel wrapper (auto CUDA/CPU)
│   ├── live_detect.py       # Live dual-model detection + Bangla TTS announcer
│   ├── utils.py             # Box geometry, proximity zone helpers
│   ├── config.py            # Model path resolution
│   ├── best.pt              # Custom-trained currency model
│   └── yolo11n.pt           # Pretrained YOLOv11 nano
└── routes/
    └── detect_route.py      # Detection API endpoints

API overview

Method	Endpoint	Description
GET	/detect/live	Start live webcam detection with voice announcements
POST	/detect/image	Detect objects in an uploaded image (single model)
POST	/detect/image/dual	Detect with both models simultaneously
POST	/detect/image/annotated	Return the image with bounding boxes drawn
GET	/detect/models	List loaded models
GET	/detect/classes	List detectable classes
GET	/detect/health	Health check

Getting started

Prerequisites: Python 3.11+, a webcam (for live mode), and an internet connection (Google TTS).

git clone https://github.com/IRIvey/ALOK.git
cd ALOK/backend

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

pip install -r requirements.txt

# Run the API server
uvicorn main:app --reload

# Interactive docs at http://127.0.0.1:8000/docs

To launch live detection with Bangla voice announcements, hit GET /detect/live (or run python detection/live_detect.py directly). Press q to quit the live window.

The custom currency model

Off-the-shelf models don't know Bangladeshi Taka notes, so we built our own dataset of currency images and trained a custom YOLO model (best.pt) to recognize denominations (e.g. ৳100, ৳500). At runtime it runs side-by-side with pretrained YOLOv11 rather than replacing it, so general object awareness is never lost.

Roadmap

• Text reading (OCR) — Bangla + English document reading with Tesseract and EasyOCR, read aloud via TTS
• Hands-free Bangla voice commands for fully non-visual interaction
• Offline TTS for low-connectivity areas
• Mobile app targeting standard smartphones
• Expanded currency dataset and scene description

Team — pip_install_chaos

Name	GitHub
Adrita Ahsan	@adrita06
Ayesha Mashiat	@ayeshamashiat
Israt Risha Ivey	@IRIvey
Nishat Tasneem Oishee	@Oishee53

Built at the SOLVIO AI Hackathon 2025 (organized by Sheba Platform, Bangladesh's largest AI hackathon), where ALOK placed in the Top 100 teams.