Panda

💡 Inspiration

Science students, early-career researchers, and lab technicians spend excessive hours manually drafting, verifying, and formatting laboratory Standard Operating Procedures (SOPs). Not only is this manual process tedious, but a single human error in noting chemical risks can lead to hazardous lab accidents. We wanted to bring "Suvidha" (Convenience and Safety) to the scientific community by building an autonomous assistant that structures workflows while strictly guarding against hazards.

⚙️ What it does

LabMate AI is a domain-specific, multi-agent AI research assistant. You provide it with a raw scientific research paper, and it orchestrates three internal agents to process it:

1. Safety Agent: Scans for hazardous chemicals, required PPE, and critical warnings.
2. Evidence Agent: Extracts key scientific claims and traces them back to the source text.
3. Planning Agent: Generates a strict, step-by-step laboratory workflow.

The Killer Feature (Safety Gate): If the AI detects highly toxic chemicals (e.g., Dichloromethane) or receives insufficient data (e.g., just typing "what enzyme"), it triggers a "RESTRICTED" Safety Gate. It halts the detailed protocol generation, preventing AI hallucinations and ensuring enterprise-grade safety.

🛠️ How we built it

• Frontend: Built using React (Vite) and Tailwind CSS for a clean, modern, and accessible UI, featuring a custom Dark/Light mode and a responsive Safety Dashboard.
• Backend: Powered by Node.js and Express.js to securely route API calls.
• AI Engine: Integrated Google's latest Gemini 3.6-flash via the @google/genai SDK. We utilized advanced Prompt Engineering to enforce strict JSON Schema outputs, ensuring the UI receives structured data (arrays, objects) instead of unpredictable text.

🚧 Challenges we ran into

Our biggest challenge was "Agentic Hallucination". Initially, if a user typed a short, out-of-context query (like "what enzyme"), the AI would forcefully invent a highly toxic, hallucinated protocol just to fill the JSON schema.

• The Fix: We engineered a "Triage & Verification Gate". We added input validation constraints so the AI evaluates the text length and context first. If data is insufficient, it safely rejects the prompt instead of hallucinating.

🎓 What we learned

We learned that building enterprise AI isn't just about API calls; it's about guardrails. We mastered multi-agent prompt orchestration, forcing the AI to reason about safety before generating actionable execution steps.

🚀 What's next for LabMate AI

We plan to deploy the backend on Google Cloud Run, integrate a live database (Firestore) for session history, and add external tool-use (fetching live MSDS chemical reports directly from the web).

Built With

• ai-agents
• express.js
• google-cloud
• google-gemini
• node.js
• prompt-engineering
• react
• tailwind-css
• vite