BuildSafeCopilot

• 

  BuildSafeCopilot

💡 Inspiration

Construction companies lose hours every week searching through safety manuals, policies, and outdated spreadsheets. On job sites, compliance questions—like fall protection rules or PPE requirements—often go unanswered until incidents occur. We wanted to build a conversational AI that turns those complex regulations into instant, actionable insights.

⚙️ What it does

BuildSafe Copilot is an AI-powered compliance assistant for construction and field operations. It: Answers OH&S and company policy questions with cited, grounded responses. Generates Job Safety Analysis (JSA) PDFs automatically based on the task, trade, and conditions. Creates audit packs with linked evidence and documents for site inspections. Combines Elastic hybrid search (text + vector) with Google Cloud Vertex AI reasoning to provide accurate, explainable answers in seconds.

🏗️ How we built it

1. Elastic Cloud (Serverless) stores all safety manuals, training docs, and incident summaries with a hybrid index (BM25 + vector embeddings).
2. Vertex AI text-embedding-004 generates dense vectors for semantic retrieval.
3. FastAPI backend merges BM25 + k-NN results, re-ranks them using Vertex AI Ranking, and feeds top chunks into Gemini 1.5 for grounded generation.
4. Next.js frontend provides a chat interface with citations and an Actions panel for JSA and audit pack generation.
5. Generated PDFs and ZIPs are stored in Google Cloud Storage, with signed links returned to users.

🧱 Challenges we ran into

Integrating Elastic’s vector search with Vertex AI Ranking smoothly required careful data formatting and batching. Handling multilingual content (EN/FR) in regulations while keeping embeddings consistent. Managing authentication and credentials across GCP, Elastic, and the local dev environment in Cursor. Keeping the demo lightweight enough to deploy on Cloud Run while preserving latency under 2 seconds.

🏆 Accomplishments that we're proud of

Delivered a fully working Elastic + Vertex AI grounded RAG pipeline within days. Built a real-time JSA generator that outputs OSHA-compliant PDFs. Achieved instant hybrid search (< 1 second average retrieval) on synthetic policy data. Created a clean, production-ready Next.js UI that mirrors enterprise safety dashboards.

📚 What we learned

How to use Elastic Serverless for scalable vector search integrated with Google Cloud AI services. The power of grounded generation Gemini performs best when paired with verified Elastic context. That small, domain-specific datasets can outperform large general LLM contexts when curated and indexed well. How to orchestrate AI agents and document workflows inside a regulated, safety-critical domain.

🚀 What's next for BuildSafe Copilot

Add multilingual support (English/French/Spanish) for cross-border crews. Integrate with Wearable IoT Sensors and site logs to provide proactive hazard alerts. Expand to Energy and Manufacturing compliance use cases. Release as a freemium SaaS platform for contractors and SMBs with fine-tuned models per trade.

Submit to Google Cloud Marketplace as an Elastic + Vertex AI template for compliance automation.

Built With

• elastic
• google-cloud
• javascript
• python
• react
• typescript
• vertexai