Project Description
I need an AI-driven chatbot that can hold text-based conversations with patients and site visitors and reliably answer general health questions around the clock. The focus is on everyday wellness topics—nutrition, exercise, sleep, basic symptom understanding—not on diagnosing or prescribing, so the language model must be tuned for safe, informational responses and include the appropriate disclaimers.
Here’s what I’m looking for in the finished product:
• A web-embeddable chat widget (React or plain JS is fine) that launches instantly on page load.
• A backend powered by a large language model (OpenAI, Azure OpenAI, or a comparable on-prem alternative) with fine-tuning or prompt-engineering that keeps every answer within the boundaries of general health advice.
• Context management so each user session feels coherent, yet no personal health data is stored after the chat ends. HIPAA compliance is essential for any transient logging you do during processing.
• An admin dashboard where I can review conversations, flag problematic answers, and update safety prompts without needing to touch code.
• Clear deployment instructions (Docker or serverless are both acceptable) so I can move the bot from your test environment to my production server with minimal friction.
Success Criteria
1. A user can ask any common health question (e.g., “How much water should I drink daily?”) and receive a concise, lay-friendly reply in under two seconds.
2. The response always includes a disclaimer to “consult a qualified healthcare professional for personal advice.”
3. The system passes a short test set of 50 sample questions I will provide, maintaining factual accuracy above 90% and never generating restricted medical diagnoses or prescription instructions.
If you have previous experience building safe, domain-specific chatbots or have implemented guardrails for medical content, your approach will stand out. Please outline the stack you’d choose, how you will handle content filtering, and links to any similar live projects.