Salamtak سلامتك

Salamtak سلامتك

Checks one site photo for safety hazards and gives a Safe, Needs attention or Unsafe verdict with short alerts in seven worker languages. Runs on vision models and plain rules, no LLM. Suits site safety teams.

Made by plitip on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Salamtak looks at one photo of a construction site and finds hazards in it. It then gives a verdict of Safe, Needs attention or Unsafe, marks the problems on the picture, and writes short alerts in the languages workers speak. The Space was built by a hackathon team in Doha.

Who it suits: It suits site safety officers, supervisors and contractors who want a quick second look at a photo, and developers who want to study a small rules-based safety pipeline. It is advisory only.

Industry | Who uses it | How they use it | Example
Construction | Safety officers | Check a site photo for missing helmets and other hazards | Upload a photo of a crew and read the verdict
Heavy industry | Supervisors | Alert mixed-language crews in their own language | Show the Arabic or Hindi alert to a team
Home safety | Parents and caregivers | Use Home mode to spot floor cords, wet floors and reachable knives or chemicals | Check a kitchen photo
Software | Developers | Read a pipeline of detectors plus a rules engine | Open the safesite folder in the repository

You can skip it if you need a certified safety audit, since no alert does not mean no hazard.

Where to use it:

• A web browser: open the Space page on Hugging Face, then upload a photo or take a webcam snapshot
• A webcam in the Live tab, for about one frame per second
• Windows, macOS or Linux: run a copy with Python 3.11

Use case | Who | What they get back
Check a photo | Safety officers | Hazards with reasons, a verdict and an annotated picture
Watch a webcam | Site teams | Hazards that show once seen in 2 of 3 frames
Alert the crew | Supervisors | Short alerts in English, Arabic, Hindi, Bengali, Nepali, Urdu or Malayalam
Check a home | Families | A Home mode check of floors and reachable dangers

How to set it up: The online page needs nothing. A local copy needs more, including system packages.

• A web browser for the online demo
• Python 3.11 for a local copy
• The Python packages listed in requirements.txt
• The system packages listed in packages.txt, installed with apt-get on Linux

1. To use it online, open the Space page on Hugging Face.
2. To run a copy, install the Hugging Face hub package.
3. Download the Space into a local folder.
4. Move into the folder and install Gradio 6.31.0 and the listed packages.
5. Start the app.

pip install -U huggingface_hub
hf download plitip/team-qh-091-safesite-app --repo-type space --local-dir team-qh-091-safesite-app
cd team-qh-091-safesite-app
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: Open the Photo tab, add a photo, and optionally set the photo time, a WBGT heat reading, whether it is outdoors, the alert language and Site or Home mode. Read the verdict and the annotated picture. The Live tab does the same on a webcam, and a demo button plays a short sample clip if you have no camera.

python app.py

Tool | What it does
Photo tab | Checks one uploaded or webcam photo and returns hazards, a verdict and alerts
Live tab | Checks a webcam stream about once a second and shows hazards seen in 2 of 3 frames
Play demo site camera | Runs a short sample clip through the same detector and rules

Example: You upload a photo of a roadside crew at noon, enter a high WBGT reading and pick English. The page marks workers without hard hats, flags the midday heat rules and missing shade or water, and shows an Unsafe verdict with a stop-work alert.

For newcomers and experts: If you are new to this, open the first example in the app and read how each hazard on the picture ties to a short reason. Then try your own photo and look at the verdict.

If you know the field, no language model is involved. A fine-tuned YOLO26n detector with 11 classes finds people, helmets, vests and equipment. YOLOE-26 text prompts find extra objects with no training. A plain Python rules engine, which includes Qatar heat rules, turns detections into hazards. Prompts ship only if they pass an image-level bench for hit rate and false alarms.

You add a photo, time, WBGT and language
A YOLO26n detector finds people and gear
YOLOE-26 prompts find extra objects
A rules engine lists hazards with reasons
You get a verdict, markup and alerts

Good to know: A photo with no people or equipment is marked Not assessed, never Safe. A few rules cover things that no shipped model can detect yet: fire work, missing extinguishers, floor openings and trenches. Each one appears under the verdict as not checked. The crane-hook check rarely fires because the training data on it is thin.

Title: Site mode hazard checks
Hazard | Severity
Missing helmet | High
Person inside a machinery or vehicle zone | High
Worker lying on the ground | High
Person under a hanging load | High
Work at height with no fall protection | High
Machinery beside a power pole | High
Outdoor work during the midday ban | High
Heat reading above the WBGT limit | High
Missing vest near a vehicle | Medium
Exposed rebar with workers near | Medium
No shade or water outdoors | Medium
Cord or hose across a walkway | Medium

Title: Hazards caught on the Qatar test set
Version 1: 25 percent
Shipped version 2n: 62 percent

• License: agpl-3
• Cost to run: none named
• Data it reaches: the photos and webcam frames you give it, plus the time and WBGT you enter
• Not for: certified safety audits, since it is advisory only

On Hugging Face, 13 November 2026

• Likes: 2
• Built with: gradio
• License: agpl-3.0

Links

• Space: https://huggingface.co/spaces/plitip/team-qh-091-safesite-app
• Files: https://huggingface.co/spaces/plitip/team-qh-091-safesite-app/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.