Quick thoughts on building AI systems, shipping to production, and what actually works.
Coordinating 44 tools across 6 MCP servers taught me more about failure handling than any tutorial. I spent two weeks debugging crashes before realizing it wasn't the LLM—it was how agents handled tool timeouts.
Read more →Integrated OpenAI API and LangChain to automate data extraction and classification at University at Buffalo. The trick was knowing when to cache responses and when real-time API calls actually mattered.
Read more →Moved from blocking operations to Celery-based async task queues at Roche, handling high-volume patient data pipelines. Response times dropped 75% after proper PostgreSQL indexing and background job architecture.
Read more →Set up CI/CD pipelines with GitHub Actions and Docker at UB, reducing deployment time from manual releases to automated pushes. The infrastructure setup with Terraform made AWS deployments reproducible across environments.
Read more →