Kevin Madura is a director at AlixPartners, where he advises companies, lawyers, regulators, and courts on cybersecurity, digital assets, and enterprise AI. His work translates complicated documents and professional judgment into inspectable, measurable language-model systems.
In 2009, Madura discovered a cross-site request forgery vulnerability in Moodle. He earned a computer science degree from the University of Maryland and a master’s degree from Georgetown University, then worked at IBM as a cybersecurity principal supporting the U.S. Department of Defense and Department of Homeland Security.
By 2019, he was a vice president in AlixPartners’ cybersecurity practice and co-authoring research on cybersecurity risk in mergers and acquisitions. His practice expanded into cryptocurrency investigations, blockchain infrastructure, digital-asset disputes, and expert testimony in federal proceedings and international arbitration. He also co-authored an analysis of the Poly Network theft.
Madura subsequently helped develop an internal generative-AI platform at AlixPartners. He treats automation as a way for experienced professionals to examine entire collections of contracts and investigative records, instead of restricting their judgment to whichever documents limited time allows them to review.
AI applications as modular software: Madura builds language-model applications as typed, testable Python programs using DSPy signatures, Pydantic schemas, adapters, and explicit evaluation metrics. His AI Engineer workshop repository demonstrates multimodal ingestion, structured extraction, model routing, token tracking, and reusable optimization. He considers DSPy’s programming model more fundamental than its prompt optimizers.
Structured extraction with accountable review: For credit agreements and other complex documents, Madura treats the extraction schema as the expression of business judgment. He uses token probabilities as an approximate confidence signal for human reviewers and classifies document pages to identify section boundaries, distinguishing experimental shortcuts from production-grade validation.
Measured optimization: Using DSPy and GEPA, Madura reported a 20-percentage-point improvement on a specific structured-extraction task. He uses task-level metrics to inspect failures and assess whether smaller, cheaper models can achieve acceptable performance.
Recursive language models: Madura embeds models inside programming environments that inspect dataframes, navigate long documents, delegate model calls, and audit source code. His benchmark harness evaluates recursive DSPy systems on data-analysis tasks.
From searching contracts to inspecting optimized modules, Kevin Madura shows how typed signatures, ordinary Python control flow, and measurable feedback fit together in DSPy.
Classification, retrieval, and structured extraction show how AI can change consulting work—and why enterprise value still depends on schemas, validation, and reliable delivery.
Kevin Madura explains how recursive language models move large inputs into a programmable REPL, where a model can inspect variables, write code, delegate focused work, and return only the results that matter.
An RLM keeps large inputs as symbolic objects in a persistent REPL, allowing generated code to inspect and transform them without placing the entire working set in the root model’s context.
The best candidates have large or dense inputs, natural decomposition, or useful deterministic computation. Small, latency-sensitive tasks may not justify the extra execution and model calls.
The 12-numbers-in-30,000-tokens example captures the core transformation: let the model select regex and Python, then let deterministic code perform extraction and arithmetic.
Typed inputs, outputs, iteration limits, traces, and schemas create a controlled shell, but they do not prove analytical correctness or guarantee good stopping behavior.
The longer-term bet is that post-training models for RLM-style execution will improve decomposition, code generation, delegation, and context selection.
