Suraksha Mcp

Report server health and whether the model inference backend is reachable.

Analyse a construction-site image (photo or CCTV frame) and detect PPE / hazard conditions using a trained multi-label classifier (10 classes: Hardhat, Mask, Safety Vest and their NO- variants, Person, Safety Cone, machinery, vehicle). Reports which conditions are present in the whole frame — it does NOT localise individual people (no per-worker bounding boxes), so treat results as frame-level, not per-person. Returns structured violations with severity so an agent can decide whether to alert, escalate, or log.

Detect real-time struck-by / run-over hazards — a worker dangerously close to moving machinery or a vehicle — from a site image. This is the DYNAMIC-hazard case where an immediate spoken warning in the worker's language can prevent an incident (cf. EMESRT L7 audible-alert collision-avoidance). Vision-based and low-cost; complements UWB tag systems. Requires an object-detection backend with real per-object boxes — with the current whole-image classifier model this will not fire (no localisation available).

Identify the spoken language from a short (~5s) audio clip of a worker, using an audio classifier fine-tuned on a 10-Indian-language speech dataset. Use this to route safety alerts to each worker in their mother tongue. Returns the top language + confidence.

Turn a safety warning into SPOKEN audio in a specific Indian language (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Urdu) using AI4Bharat Indic-TTS. This is what reaches low-literacy workers who cannot read signage. Accepts either a known violation type (uses a verified canned phrase) or free text.

Generate a short SPOKEN safety briefing (toolbox talk / site induction) in a chosen Indian language for a topic (e.g. "working at height", "excavation near vehicles"). Addresses the real training gap for multilingual migrant crews — a stronger use of voice than narrating signs. Each delivered talk is logged as a training session, feeding the BRSR EI-8 training-coverage metric in generate_esg_report.

Aggregate a set of detected violations into an audit-ready, regulation-cited safety report (compliance rate, breakdown by type/severity, cited clauses, corrective actions). Also appends each violation to the incident log resource.

End-to-end safety audit over a batch of site images (and optional worker audio): detects PPE violations AND real-time proximity/struck-by hazards, identifies each worker's language, and generates spoken multilingual alerts for non-critical findings. Critical hazards are NOT auto-alerted — the task flags them (input_required) and returns them under pendingApproval; call generate_voice_alert to approve and fire each one. Returns a regulation-cited report. Runs as a long-running MCP Task with streaming progress.

Generate a BRSR Principle 3 (SEBI ESG disclosure) aligned safety report from the persistent incident/training/alert database: leading indicators (detected violations, trend), training coverage, measures taken, plus LTIFR and other lagging indicators when HR data is supplied. Each BRSR field is labelled measured / provided / requires-input. Use this to help an MSME contractor evidence safety compliance to BRSR-reporting clients.

One-time setup: store a site's own Twilio credentials for WhatsApp alerts, so that site's messages send from its own Twilio account instead of the server default. Call this once per site, then use send_whatsapp_alert normally — it will resolve the right credentials by siteId automatically. The auth token is stored server-side only and is never returned by any tool, including this one.

Send a safety alert to a phone over WhatsApp (Twilio). Pair with generate_voice_alert: pass its audioUrl as mediaUrl so a low-literacy worker RECEIVES the spoken alert in their language on their own phone. Resolves credentials in order: the site's own config (set via configure_alert_channel), then the server default (TWILIO_* env vars). With neither configured, returns a dry-run payload instead of sending (safe for demos). Sends are logged as safety measures for ESG reporting.