A healthcare agentic AI assistant that automates front-desk operations by communicating with patients through messaging platforms (WhatsApp and others). It handles scheduling, routine questions, appointment reminders, follow-ups, and feedback collection. It is integrated with 20+ clinic management services.
- Revived this project by rewriting the legacy solution and architecture into a new modular one with frameworks for AI (Python, TypeScript, architecture, system analysis, LangGraph, LangChain, JavaScript).
- Made clinic documents usable for the AI by ingesting them with LlamaIndex into PostgreSQL plus PGVector (LlamaIndex, PostgreSQL, PGVector, Vector DBs, RAG, Python, BackEnd, SQL, ETL).
- Made an AI Agent with RAG in production by writing agentic pipelines and implementing tool calling with LangChain and LangGraph (LangChain, LangGraph, AI Agents, tool calling, RAG, workflow, Python).
- Added observability with LangSmith on retrieval functions and reranking functions so each turn showed what was retrieved, what was reranked, and what the agent answered (LangSmith, retrieval, reranking, LLMOps, LangGraph, LangChain).
- Increased the quality of retrieval by using advanced techniques (multi-query, hierarchical retrieval, reranking, BM25, RRF, different chunking strategies, structure-aware chunking, MMR).
- Kept clinic and patient data fully private by running the local LLM of the AI agent self-hosted on a cloud server. This way the data was not exposed to ChatGPT (OpenAI), Claude (Anthropic) and Gemini (Google). (Cloud, AWS, BackEnd, Azure, GCP, private LLM, self-hosted, data privacy).
- Solved critical prompt-injection vulnerabilities and credential exposure risks by adding a prompt guard and implementing MCP as a secure middleware layer for calling internal microservices and third-party clinic management software (MCP, Prompt, architecture, API, Microservices, FastAPI, FastMCP).
- Scaled adoption among clinic clients across multiple platforms by connecting AI to messaging apps (WhatsApp, Viber and others) (BackEnd, webhooks, Python).
- Kept the AI agents online while we shipped new features by implementing a microservice architecture and deployments with Kubernetes. Rolling deploys ran those same services in production with no clinic downtime. The testing environment was in Docker, Docker Compose (Kubernetes, Docker, Microservices, Pods, Docker Compose, QA, testing).
- Increased the speed of RAG by using Redis for caching frequently asked questions across multiple users (Redis, BackEnd, embeddings).
- Kept the AI chat with customers responsive at all times by executing agentic actions asynchronously via RabbitMQ (event-driven architecture, task queuing, async, execution, async messaging).
- Improved the UI/UX of the web app for clinic managers by showing how the AI thinks and where it retrieved its information (FullStack, FrontEnd, Next.js, BackEnd, React, React.js, JavaScript, Python, UI/UX, SSE, React Hooks).
- Python
- TypeScript
- architecture
- system analysis
- LangGraph
- LangChain
- JavaScript
- LlamaIndex
- PostgreSQL
- PGVector
- Vector DBs
- RAG
- BackEnd
- SQL
- ETL
- AI Agents
- tool calling
- workflow
- LangSmith
- retrieval
- reranking
- LLMOps
- multi-query
- hierarchical retrieval
- BM25
- RRF
- structure-aware chunking
- MMR
- Cloud
- AWS
- Azure
- GCP
- private LLM
- self-hosted
- data privacy
- MCP
- Prompt
- API
- Microservices
- FastAPI
- FastMCP
- webhooks
- Kubernetes
- Docker
- Pods
- Docker Compose
- QA
- testing
- Redis
- embeddings
- event-driven architecture
- task queuing
- async
- execution
- async messaging
- FullStack
- FrontEnd
- Next.js
- React
- React.js
- UI/UX
- SSE
- React Hooks