Sano: Personal Health Assistant Chatbot

Problem

The project started from medication non-adherence: about half of prescriptions are not taken as prescribed, which drives chronic-disease treatment failures. The goal was an assistant that acts like a caregiver: remind, explain the medicine, and help find it.

Approach

Messages arrive from Telegram through a webhook to the Sano bot server. RASA NLU (DIETClassifier with spaCy multilingual entity extraction) detects the intent and entities, so "remind me to take a multivitamin every day at 8 am" becomes a reminder request with the medicine and schedule. A fine-tuned MobileNetV2 identifies drugs from photos, and TF-IDF with cosine similarity over medicine descriptions recommends alternatives.

How It Works

1. 1
   MessageThe user chats with Sano on Telegram; a webhook forwards each message to the bot server.
2. 2
   UnderstandRASA NLU classifies the intent (e.g., set a reminder) and extracts entities such as the medicine and time.
3. 3
   IdentifyA fine-tuned MobileNetV2 recognizes a drug from a photo of the pill or package.
4. 4
   RecommendTF-IDF vectors and cosine similarity over medicine descriptions suggest similar drugs.
5. 5
   Remind & locateReminders are stored and sent on schedule, and Sano points to a nearby pharmacy.

What I Built

• The Telegram bot server, webhook handling, and the reminder database.
• The RASA NLU pipeline (DIETClassifier, spaCy multilingual entity extraction) for reminders, drug questions, and recommendations.
• A fine-tuned MobileNetV2 image classifier for drug identification.
• A similar-drug recommender using TF-IDF and cosine similarity on medicine descriptions.

Impact

One chat interface covers four caregiver tasks: medication reminders, drug identification from text or a photo, similar-drug recommendations when stock runs out, and directions to a nearby pharmacy.