Isaac Miller is a co-founder of cmpnd and a lead maintainer of DSPy, the open-source framework for building AI applications as modular, measurable programs. He focuses on keeping software adaptable as models, prompts, tools, and reasoning strategies change while holding its behavior accountable to concrete product requirements.
From software engineering to AI infrastructure
Miller studied computer science at Northwestern University, founded its AI Club, and worked as a graduate research assistant investigating how language-model systems could improve in production. He previously developed mobile software at Drift Net Security and Lyft and worked on backend infrastructure at Robinhood.
By August 2024, he was publicly advocating compound AI systems built around explicit objectives and measurable results. His case for DSPy criticized the assumption that chaining model calls produces dependable software and acknowledged the framework’s own shortcomings in reliability, documentation, and developer experience.
- Stable task contracts: Miller separates an application’s inputs, outputs, constraints, and evaluations from its models, prompts, tools, and execution strategies. That architecture lets teams change implementations, test emerging techniques, and compare costs without rewriting surrounding product logic.
- Evaluation-grounded optimization: He argues that new models and agent harnesses must demonstrate improvements against the specific task they serve. His emerging work on qualitative learning explores whether production traces, user actions, product analytics, and written feedback can help systems refine their evaluations over time; it remains a research direction, not an established capability.
- Recursive language models in DSPy: Miller helped integrate a long-context technique developed by other researchers into DSPy, preserving compatibility with existing task signatures. He announced the dspy.RLM release in January 2026, making the approach available as a replaceable component inside existing applications.
- Practical deployment with dspy-cli: With Drew Breunig, Miller developed dspy-cli, which turns DSPy programs into deployable services with HTTP endpoints, OpenAPI specifications, Docker configuration, and optional Model Context Protocol support.
- Context-aware Discord moderation: His DSPy community moderation system evaluates a message’s primary intent, author, and channel to distinguish genuine recruitment posts from legitimate technical discussion, while limiting permissions and keeping moderators informed.
Miller also implemented DSPy.Image and contributed ReActV2, including structured interaction history, native tool calling, parallel tool calls, and improved tracking of tool results. His AI Engineer World’s Fair appearance crystallized his view that even increasingly capable models need application-specific context, explicit evaluation, and learning grounded in real production behavior.
