Arogya AI
A conversational AI prototype connecting Rasa intent classification and Python actions to a browser chat interface through a REST webhook.
My focusConversational AI & web integration
Educational prototype, not a clinically validated diagnostic tool. The current web client targets local Rasa; multilingual WhatsApp/SMS delivery is a broader project goal, not verified in this source.
- declared intents
- 9
- custom actions
- 2
- coded disease-information topics
- 10
Counts describe the reviewed source files, not medical accuracy, real-world usage, or evaluated language coverage.
Overview
Arogya AI is a Rasa-based chatbot prototype for health-information conversations. It brings together NLU configuration, dialogue policies, Python actions, and a lightweight browser interface.
The problem
A conversational application needs more than a text box: it must classify a request, select an appropriate workflow, and return a response from application logic. For health-related content, the boundary between an educational prototype and a validated medical product is particularly important.
Approach
Define conversational intents in the Rasa domain, train the NLU pipeline, and route the two health-query workflows to custom Python actions. Keep the browser interface separate from dialogue and response logic, connected through Rasa’s REST webhook.
System design
Browser message → Rasa REST webhook → NLU intent/entity processing → dialogue policy → Python custom action → response. The configuration uses a DIET classifier, count-vector features, entity synonyms, and rule-based plus learned dialogue policies.
Implementation
The domain declares 9 intents and a disease entity/slot. Two custom actions handle disease-information lookup and symptom-pattern matching. The disease-information action contains 10 coded topics and synonym matching. Despite its name, action_pubmed_qa currently uses a local dictionary; it does not query PubMed or run a retrieval-augmented model.
Engineering tradeoffs
The current knowledge base and keyword matching make the response path easy to inspect, but limit coverage and ambiguity handling. Missing information can produce an incomplete answer, so clinical validation, reviewed content, and robust refusal/escalation behavior would be prerequisites for real healthcare use.
Results & evaluation
The repository demonstrates 9 declared intents, 2 custom Python actions, and a web client posting to a local Rasa endpoint. No clinical accuracy score, production usage count, or measured response-time claim is made. The repository description mentions multilingual WhatsApp/SMS delivery, but that deployment is not verified by the reviewed files.
What I learned
This project provides experience with dialogue modeling, custom action contracts, and frontend-to-chatbot integration. Future work could add a repeatable conversation evaluation set, reviewed knowledge sources, and production channel adapters; those are improvements to pursue, not completed achievements.