I built the Multi-Modal AI Safety Assistant, a Python and Streamlit project on my GitHub, to analyze images and user questions for real-world hazards such as electrocution, fire, and slip risks. The project description says it started in 2026-06 and was last updated in 2026-06. My GitHub profile also lists the project among my computer-vision and language work in github profile.
I use BLIP image captioning, described as Salesforce via HuggingFace, to turn an uploaded image into a caption. The app sends that caption together with the user's question to Llama 3.3 via Groq; the README identifies the hazard-analysis model as Llama 3.3-70B via the Groq API. Llama returns a structured hazard report that the Streamlit interface displays.
I designed the report to explain what the danger is, why it matters, and what to do about it. The README's worked example is a frayed electrical cord beside a puddle on concrete: it identifies electrocution risk from a frayed live wire in contact with standing water, explains that water conducts electricity, and advises avoiding the cord and water, cutting power at the breaker, and calling a licensed electrician before using the area again.
The documented project structure assigns app.py to the Streamlit UI, src/image_capture.py to BLIP captioning, and src/hazard_analysis.py to Groq/Llama hazard analysis. requirements.txt lists dependencies, and the root .env file holds the API key and is not committed. The stated language requirement is Python 3.10+.
git clone https://github.com/your-username/ai-safety-assistant.git
cd ai-safety-assistant
pip install -r requirements.txt
# Create a root .env file containing:
GROQ_API_KEY=your_groq_api_key_here
streamlit run app.py
The Quick Start clone command uses a placeholder URL, so I do not have the actual repository clone URL established here. The project description mentions fire and slip hazards, but the README's worked example covers electrocution; it gives no implementation or evaluation details for those other hazard types. The documentation provides no measured accuracy, latency, or evaluation results, and does not specify model configuration, prompts, image constraints, error handling, privacy practices, or safety validation.