Mayo Oshin

Mayo Oshin builds AI tools for extracting data from documents and teaches LLM application development. His work includes a PDF-chatbot template and Learning LangChain, coauthored with Nuno Campos and published by O’Reilly in February 2025. He also describes himself as an AI advisor and an early-stage angel investor in AI development tools and platforms.[2][4][5]

Oshin’s PDF-chatbot work was publicly identified by June 2023, when Latent Space linked to “Mayo’s PDF Chatbot” in its discussion of OpenAI’s functions update. The project also appears on his personal website and GitHub profile. Its later template serves as an accompanying example to Learning LangChain.[1][2][3][4]

The archived template separates document ingestion from question answering. It parses uploaded PDFs and stores their embeddings in Supabase. A retrieval graph decides whether a question needs document retrieval or a direct answer, then generates a response using OpenAI or another LLM provider, with references when documents are retrieved. LangChain and LangGraph provide orchestration; LangGraph’s state-machine approach makes the workflow’s steps available for visualization and debugging.[4]

The example distinguishes persistent documents from temporary conversation state: ingested documents remain in the vector database across sessions, while the demonstration UI does not preserve conversations across sessions. The repository was archived on March 27, 2026, and is retained for reference rather than actively maintained.[4]

Earlier, in the June 2023 Latent Space discussion, Oshin questioned the excitement surrounding OpenAI’s larger context window and function calling. For question answering, he argued that retrieval would remain useful, especially as embeddings became cheaper; he mentioned named-entity recognition as a possible use for the additional context. He saw function calling as a shift from approaches heavy in prompt engineering toward model fine-tuning, rather than an entirely new tool-use pattern.[1]

Oshin also asked why OpenAI might not use JSONformer-style constrained generation, while acknowledging that he did not know its implementation. His initial suggestion that constrained output eliminated prompt injection drew a correction from Simon Willison: valid JSON could still encode an unwanted function call induced by malicious input. Oshin then clarified that the boundary between user input and function calling remained vulnerable. The exchange distinguished output syntax from the safety of the action requested.[1]

In his closing remarks, Oshin expressed a clear preference for open source: “Open source a hundred percent.” He also urged OpenAI to become more open.[1]

Oshin and Campos’s Learning LangChain was published in February 2025. ACM subsequently listed both as speakers for an LLM-app TechTalk scheduled for May 8, 2025, identifying Oshin as an AI engineer and Campos as a software engineer at LangChain.[5][6]

Before this documented AI work, Oshin wrote about habits, productivity and decision-making. His essay archive lists “Emotional Intelligence: The Difference Between Good and Great Leadership” and “5 Differences Between Amateurs and Professionals” under 2017. This earlier work adds a writing dimension to his career alongside his later technical projects and teaching.[7][2][5]