govardhansatya/surakshamcp

SurakshaMCP is a construction safety tool built as an MCP server, meaning it exposes its capabilities as structured tools that any AI assistant supporting the Model Context Protocol, such as Claude or ChatGPT, can call and use. It was built for a hackathon focused on occupational and construction safety, and it targets a specific real problem: construction sites in India, especially small contractors, often have low-literacy, multilingual, migrant workforces that English-only safety dashboards and signage never reach. The README cites a study estimating around 38 fatal construction accidents a day in India, with falls from height and structural collapses making up most documented deaths. The tool works by analyzing a photo from a construction site to detect personal protective equipment and hazard violations, using a fine-tuned YOLOv8 vision model. A second model identifies which language a worker is speaking, so that safety alerts can then be spoken back to them in their own language, covering ten Indian languages through a text-to-speech system called AI4Bharat Indic-TTS. The project frames this voice and language capability as especially useful for real-time hazards, like a worker straying near moving machinery, and for multilingual safety briefings with migrant crews, rather than just reading out static signs. Every detection and alert is also logged to a local database and can be rolled up into a safety disclosure report aligned with an Indian regulatory framework called BRSR, which is mandatory reporting for India's largest listed companies and increasingly expected from their supply chains. Alerts can additionally be sent to a site foreman's phone over WhatsApp using Twilio. Architecturally, the project is split into two services: an MCP server written in TypeScript using a framework called NitroStack, which is the main deliverable, and a separate Python FastAPI service that runs the actual PPE detection, language identification, and text-to-speech models. The MCP server talks to the inference service over HTTPS. A live demo endpoint is provided that any MCP client can connect to directly. To run it locally, you need Node.js 20.18.1 or newer, Python 3.10 or newer, and either run both services manually or start them together with Docker Compose. The underlying datasets and models are drawn from free and openly licensed sources, though the README notes the vision model's own license is AGPL-3.0.