Inspiration
Fever is one of the most common symptoms worldwide, yet people often struggle to understand what level of care they need. Some ignore warning signs, while others seek hospital visits for situations that could be handled with simple home care. We wanted to build something simple, accessible, and helpful: a friendly AI assistant that converts everyday language into medical insights, identifies risk, and shares evidence-based guidance in real time.
What it does
FeverChatBot is an AI-based medical assistant that:
✅ Understands free-text symptom descriptions ✅ Extracts structured medical data like temperature, duration, and symptoms ✅ Predicts fever risk level using a machine-learning classifier ✅ Retrieves reliable care tips using a lightweight Retrieval-Augmented Generation (RAG) system ✅ Presents everything inside a clean chatbot interface
Example: User → “I have 102 fever for 3 days with vomiting and body pain” Bot →
{ "temperature": 102, "duration_days": 3, "symptoms": ["vomiting", "body pain"], "red_flags": false }
Then FeverChatBot shares risk level and precise guidance such as hydration reminders, safe fever management, red-flag monitoring, and when to consider medical attention.
How we built it
Fine-tuned LLaMA-3-8B using LoRA (Unsloth) to enforce medical reasoning + strict JSON output
Created a custom fever-focused dataset and merged it with medical conversational data
Trained a Random Forest classifier to estimate severity based on extracted features
Implemented a lightweight RAG system using TF-IDF + cosine similarity to pull medically grounded advice
Designed a Streamlit chatbot UI that runs locally and doesn’t need commercial APIs
Challenges we ran into
Medical LLMs can hallucinate unpredictably, so enforcing JSON structure and safety rules required careful prompt design
Running an 8B LLM locally required optimization (4-bit quantization and LoRA adapters)
Symptoms can be vague or incomplete, so the system needed to gracefully handle missing or partial information
The ML model and LLM needed to agree on feature formats, requiring strict data mapping
Accomplishments that we're proud of
Built a working medical chatbot that runs offline, without paid APIs
Successfully fine-tuned a large LLM to extract clean JSON consistently
Combined LLMs + ML + RAG into a single unified pipeline
Real-time interaction with quick inference
Simple UI that anyone can use
What we learned
Prompt engineering dramatically affects reliability in healthcare-style chatbots
LoRA fine-tuning makes large models trainable even on small hardware
RAG helps ground answers to real medical guidelines
Combining classical ML with LLMs boosts accuracy and stability
Good UX matters — users need clarity, not technical output
What's next for FeverChatBot
Expand coverage to more fever-associated conditions and regional illnesses
Add multilingual communication (Hindi, Tamil, Bengali, etc.)
Deploy as a web app, WhatsApp bot, or offline mobile app
Integrate real-time outbreak awareness and regional health tips
Add continuous chat memory and follow-up question analysis
Built With
- machine-learning
- ollama
- python
- rag