Yogendra Miraje is a principal AI engineer at FactSet building governable financial-research agents: systems that turn institutional expertise into controlled, reusable workflows. His work traces a progression from financial-data infrastructure to autonomous research assistants whose planning, tools, evaluations, and operating boundaries remain accessible to human oversight.
Miraje earned an engineering degree in India and a master’s degree in computer science from Northeastern University. At Truvalue Labs, he developed foundational back-end and machine-learning technologies before FactSet acquired the company in 2020.
At FactSet, he helped integrate AI into its financial-data ecosystem. His writing on conversational access to financial data examines how language models become useful to investment professionals when grounded in trustworthy information and actual research workflows.
- Controllable agent planning. Miraje distinguishes fixed workflows executed by agents from dynamic workflows agents plan themselves. His planning architecture adapts the externally developed LLMCompiler design into four LangGraph components: blueprint generator, planner, executor, and joiner. Natural-language blueprints decompose objectives, restrict available tools, and let domain experts inspect behavior. Preparing for an NVIDIA earnings call, for example, combines previous-call analysis, financial-data retrieval, question development, and report generation.
- Agent skills as product features. Miraje replaced proprietary blueprints with the open Agent Skills standard, locating task-specific business logic in reusable skills. His skill-centric harness combines a skill registry, system prompt, file-reading tool, and execution loop, with sandboxed scripts where necessary. Prompts define the agent’s role, tools establish its available connections, and skills determine how particular jobs get done.
- Intent-based skill routing. Miraje organizes skills around customer objectives, such as earnings preparation, instead of underlying datasets. Carefully differentiated descriptions help agents select appropriate capabilities: an explicit PDF request activates PDF generation, while another publishing request can produce HTML. Larger skill libraries require retrieval, specialized routing models, hierarchies, and metadata filters.
- Enterprise skill governance. Miraje advocates named maintainers, human-reviewed admission, access-controlled tools, versioning, deprecation policies, and periodic audits. Because model upgrades can change how unchanged instructions are interpreted, he emphasizes component-level and end-to-end evaluations. When deterministic outcomes, tight latency, or limited budgets dominate, he favors conventional pipelines over autonomous agents.
Miraje also produces AI Blindspot, a podcast exploring emerging AI research and enterprise applications.
