Hakeema

Your Digital Health Partner for Appointments, Reports & Care, All in One Place.

Comment

Inspiration

Our inspiration is directly aligned with Google's mission: to organize information and make it universally accessible and useful. We saw a critical gap in the healthcare sector, where essential information for patients is often disorganized, inaccessible, and difficult to act upon. The stress of dealing with health issues is frequently compounded by the administrative friction of simple tasks like booking an appointment or registering at a clinic. We were inspired to apply Google's state-of-the-art ADK framework to solve this real-world problem, creating Hakeema, a platform that makes healthcare interactions simple, accessible, and useful for everyone.

What it does

Hakeema is a SaaS platform that provides clinics with a powerful, AI-driven conversational agent, built on Google's technology, to automate and enhance the patient experience. The platform serves as a 24/7 digital front door, intelligently handling crucial patient interactions.

Key functionalities include:

  1. Intelligent Appointment Booking: Users can schedule appointments using natural language. The system captures and validates details like appointment type, desired date/time, and department.

  2. On-Demand Information: Patients can instantly check visiting hours for any department, getting clear and immediate answers.

  3. Secure Test Result Inquiries: Authenticated patients can inquire about the status of their lab or imaging results in a secure manner.

  4. Streamlined Bill Payments: The platform facilitates the payment of medical bills by identifying the patient and guiding them through a simple process.

  5. Effortless New Patient Registration: New users can register with the clinic through a guided conversational flow that captures all necessary demographic and insurance information accurately.

  6. Structured Feedback Collection: The system allows patients to easily submit compliments or complaints, providing clinics with valuable, actionable feedback.

How we built it

Hakeema is built from the ground up on a Google-centric technology stack, leveraging the power and scalability of Google's AI ecosystem. The architecture is a sophisticated, multi-agent system designed for robust and modular task handling.

  1. Core Architecture with Google ADK: The system is orchestrated by a root_agent built using the Google Agent Development Kit (ADK). It triages user queries and routes them to specialized agents for greetings (greeting_agent), answering FAQs (kb_agent), or handling actions (intent_entity_agent).

  2. Intelligence Powered by Gemini: At the heart of Hakeema is Google's powerful and efficient Gemini 2.0 Flash model. Gemini's advanced reasoning capabilities are crucial for understanding user intent, carrying on natural conversations, and making decisions within the agentic framework.

  3. High-Accuracy Intent Recognition: The intent_entity_agent uses a sub-agent for intent prediction. This process leverages a Chroma vector database to perform a semantic search for the user's query against our defined intents. The results are then passed to Gemini, which uses its reasoning to make a highly accurate final prediction, ensuring the correct workflow is triggered.

  4. Dynamic Entity Collection & Validation: Once the intent is known, the intent_entity_agent masterfully conducts the conversation: a) It determines the required information (entities) for the task. b) It engages the user in a multi-turn dialogue to gather the data. c) Crucially, each piece of data is validated in real-time using specific tools (today_or_future_date, ten_digit_number, etc.) before being saved. This ensures data integrity before any action is taken.

  5. Action Execution: Once all information is collected and confirmed with the user, the action_executor_agent takes over. It calls the final tool (e.g., register_patient) to complete the task, which would integrate with a clinic's backend systems in a production environment.

Challenges we ran into

  1. Intent Ambiguity: User language can be vague. We solved this by designing a dedicated intent_prediction_agent with a strict prompt that leverages Gemini's ability to choose from a list of candidates. Forcing it to select only from the top retrieved intents—or decline if confidence is low—made the system far more reliable.

  2. Maintaining Conversational Flow: Creating a natural dialogue while methodically collecting data is complex. We crafted the INTENT_ENTITY_AGENT_PROMPT to act as a detailed "playbook" for Gemini, guiding it through the logical steps of the conversation to feel less robotic and more human.

  3. Data Validation and Integrity: In a healthcare context, data quality is non-negotiable. Our key challenge was building a flexible validation framework. By mapping entities in our configuration to specific validation tools, we created a robust gatekeeping system that ensures all data is verified before use.

Accomplishments that we're proud of

  1. We successfully built a complex, multi-agent system by harnessing the power of the Google Agent Development Kit and Gemini, demonstrating the potential of this stack for sophisticated task automation.

  2. Our two-step intent recognition process, which combines semantic search with Gemini's reasoning, achieves a very high level of accuracy.

  3. The dynamic entity collection pipeline, with its real-time validation, is a robust solution for ensuring data quality in sensitive applications.

  4. We've created a highly extensible architecture where new capabilities can be added with minimal code changes, showcasing a scalable and maintainable design.

What we learned

This hackathon was a deep dive into the practical application of Google's AI tools. We learned that a powerful model like Gemini truly shines when it's part of a well-designed agentic system. The importance of precise prompting and structured instructions cannot be overstated; it's the key to unlocking the model's full potential for reasoning and task completion. Furthermore, we validated that using retrieval-augmented generation (RAG) for intent-matching is a highly effective and scalable strategy.

What's next for Hakeema

We envision Hakeema becoming a comprehensive patient engagement platform, deeply integrated within the Google ecosystem to provide even more value.

  1. Google Wallet Integration: Allow patients to save and pay their medical bills seamlessly using Google Wallet.

  2. Google Maps Integration: After an appointment is booked, automatically offer to send directions to the clinic via Google Maps.

  3. Deeper Integration with Google Cloud: Migrate the backend infrastructure to Google Cloud for enhanced scalability, security, and the ability to leverage services like Cloud Functions for our tools and Cloud SQL for our database.

  4. Proactive Engagement: Implement the notification_agent to send proactive reminders and updates via channels like RCS Business Messaging.

  5. Voice Integration with Google Assistant: Allow users to interact with Hakeema hands-free through Google Assistant, making it even more accessible.

Built With

Share this project:

Updates