GenAI · Engineering case study
AI Agent Platform
Client: Pollstar
Intelligent, tool-connected agents for the music industry. Built for multi-step execution and production observability.
The problem & context
Pollstar, a live-entertainment industry publication, connecting language models with internal and external music APIs through a web app and agent.
Architecture
A React chat interface calls FastAPI on Azure App Service, which runs a single LangChain and LangGraph agent with memory and checkpointing. The agent reasons with GPT-5.4 and GPT-5.4 mini through Azure AI Foundry, calls tools hosted on an Azure Function App that consume the existing .NET music APIs, and uses retrieval-augmented generation over a FAISS index to read documents. Azure Cosmos DB holds conversations, agent state, and metadata, while LangSmith and Langfuse support tracing, evaluation, prompt monitoring, and cost tracking.
Illustrative system view, not a production screenshot.
Implementation & my contribution
- Built the web interface with React, Tailwind CSS, and shadcn.
- Developed backend APIs and agent workflows using LangChain and LangGraph.
- Integrated Azure AI Foundry models and Azure Cosmos DB for conversations, agent state, and metadata.
- Implemented memory, checkpointing, multi-step execution, structured outputs, retries, and validation.
- Used LangSmith for tracing and evaluation and Langfuse for prompt monitoring and cost tracking.
Outcome
Delivered tool-connected agent workflows with observable execution; optimized API performance, token usage, and response latency.