Yogendra Miraje

Yogendra Miraje, also known publicly as Yogi Miraje, introduced himself as a principal AI engineer at FactSet in his AI Engineer World's Fair 2026 talk. He argued that reusable skills become product features: they contain the procedures and business logic an agent follows, while engineers build the harness that discovers and executes them.[4][5]

Miraje's team initially used blueprints—step-by-step recipes that prevented an agent from having to rediscover an execution path for familiar work. In his 2026 talk, he explained why they adopted Anthropic's shared skills format: “there was no point in trying to maintain our own standards”. He framed the division of responsibilities simply: prompts define the agent's identity, tools define its connections, and skills explain how to complete a task.[5]

His minimal skill-aware harness uses a registry, a system prompt and a file-read tool; scripts additionally require an execution environment. The prompt contains skill names, descriptions and file paths, while full instructions load only when selected. In his NVIDIA-report demonstration, the agent reads a company-research skill, searches the web and produces Markdown, then loads a reporting skill to turn that research into HTML. Both stages run through an agent loop that appends tool interactions to the conversation and continues until the model returns its answer.[5]

Miraje treats descriptions as routing signals. His PDF-report skill activates only when the user explicitly requests PDF, distinguishing it from HTML reporting. He also recalled repeatedly refactoring a library initially organized around data categories such as estimates and fundamentals. His recommendation was to “cut by user intent and not by data model”: earnings preparation can incorporate estimate analysis, while pre-market preparation can combine news and analyst ratings.[5]

A model upgrade supplied a concrete warning about relying on unchanged instructions. Miraje reported that his agent stopped obeying skills even though their text had not changed. He attributed the failure to the new model focusing on the beginning of a skill while critical instructions appeared at the end. He urged teams to rerun evaluations whenever they change models, describing skills as behavioral contracts tied to a model. His account did not identify the models or quantify the failures.[5]

As a library grows, Miraje recommends replacing an all-in-one prompt catalog with discovery mechanisms such as embedding search or a smaller model that shortlists relevant skills. For enterprise libraries, he described automated admission gates with human review and called for named maintainers, semantic versions, deprecation warnings, changelogs and periodic audits. He also emphasized tool allowlists and access-controlled tools to preserve boundaries around what skills can do.[5]

Before his skill-centric work, Miraje explained the engineering behind natural-language access to enterprise data in “Using Large Language Models to Converse with Your Data”, published August 26, 2024, with Graham Nelson credited as a contributor. The article used FactSet Mercury's beta as application context and the query “Show me the top 50 banks by assets in California” to show how much work sits beneath a simple conversational request.[3]

Answering that query requires identifying the relevant fields and data sources, retrieving records, sorting and filtering them, and presenting the result. Miraje emphasized semantically rich metadata that connects words in the request to fields such as bank-name, bank-id, state-name and asset-value. He proposed handling retrieval outside the LLM where that could reduce generated-code errors, leaving the model to focus on data operations. He also stressed secure code execution and graceful handling of missing data or execution failures.[3]

For an analytical question such as whether large banks are more profitable than small banks, Miraje recommended a domain knowledge base to clarify the assumptions behind the answer—including what counts as a large bank and which metrics define profitability. This earlier work explains the data interpretation and execution problems behind the financial workflows he later discussed as skills.[3][5]

Earlier still, Miraje worked at Truvalue Labs, later acquired by FactSet, contributing to foundational back-end and machine-learning technologies. FactSet's undated author biography describes him as Lead Machine Learning Engineer, responsible for leading AI integration into its data ecosystem. He earned a master's degree in computer science from Northeastern University and a bachelor's degree in engineering in India.[2][3]