← All speakers

Kevin Madura

Conference affiliation: Director, Advanced Technology · AlixPartners · 2026

Kevin Madura is an AlixPartners director and expert witness who has analyzed early Bitcoin source code and transaction records for a federal jury. With Mo Bhasin, he co-led development of the firm’s internal generative AI platform so specialists could apply their judgment across an entire portfolio of contracts, even when a review budget would ordinarily cover only a fraction.

On this page

From cybersecurity to digital-asset disputes

In 2009, Moodle credited Kevin with reporting a cross-site request forgery flaw in its forum software. He holds a computer science degree from the University of Maryland and a master’s from Georgetown University. Before joining AlixPartners, he was a cybersecurity principal at IBM, where he helped the U.S. Departments of Defense and Homeland Security design and operate cybersecurity systems.

By 2019, Kevin was a vice president in AlixPartners’ cybersecurity practice and had co-authored a study of cyber risk in mergers and acquisitions. His expert witness work has included explaining how programs functioned in federal civil and criminal trials and in an international arbitration in Singapore. In 2021, he and colleagues Colin Johnson and Mark Doughty analyzed the Poly Network theft. They examined messages the attacker left on Ethereum and explained why the public trail of transactions made it difficult to spend the stolen assets undetected.

Reading a whole contract portfolio

For Kevin, the consequential choice in automated contract review is which fields a specialist asks the model to extract. An investigator and a buyer can use the same software to ask different questions of the same documents. In his work with Mo on credit agreements, those fields included the senior lender, maturity date and interest rate. To help reviewers decide what to check first, the team used the probabilities the model assigned to the tokens in each extracted value. For an interest rate, that meant scoring the words and numbers that expressed the rate. Predictable field names and punctuation played no part in that rough confidence estimate.

Kevin and Natalia Connolly built a procurement pipeline that processed more than 40,000 contract documents overnight after a merger. In their reconstruction of the case, Claude Code calculated annual spending correctly, yet its total contract value was off by about $800,000 because an amendment had lost its link to its parent agreement. The agent also gave the buyer a deadline a month late. It had found a date in the termination clause, but the deadline to prevent automatic renewal had to be calculated from the renewal clause. Kevin favored encoding deal-specific rules in database views and reusable tools. A later analyst could retrieve the relevant rate-card page and check the terms before entering negotiations.

What Kevin specifies and what the model decides

Kevin values DSPy’s programming model because he can specify the inputs and outputs of a step in Python, then change the model or its instructions while keeping the surrounding program intact. His workshop code shows how he routes different kinds of documents to different models and tracks their token use. His experiments explore how much of the work between input and answer he can leave to the model.

  • Finding document boundaries. In Kevin’s prototype for separating agreements from their schedules, a fast model labels page images. A stronger model can then inspect the original pages where the labels suggest a transition, such as a signature page followed by a rate table. A fixed set of labels makes the results easier to use in code, at the cost of assuming the developer knows what kinds of pages to expect.
  • Testing a cheaper model. In a financial-news extraction experiment, Kevin used the GEPA optimizer to revise instructions in response to errors, including confusion between company names and products. With GPT-4.1-mini doing the extraction throughout, exact matches rose from 32.07% with a direct API call to 53.84% with DSPy and GEPA. Changing how the program presented the schema brought the optimized result to 54.43%. For Kevin, the practical question is whether a smaller model can meet the accuracy a particular job requires. He calls the business value generated per dollar spent on tokens token capital efficiency.
  • Giving a model direct access to data. Kevin proposed a DSPy protocol that lets a program pass objects such as dataframes into a model’s Python environment. In his cohort-retention experiment, a recursive language model wrote and ran code against three tables of simulated data. The rows stayed outside its context window. Kevin specified the results he wanted, including a churn rate, and the model chose the calculations and when to return an answer. He also tried the approach in a source-code security audit. Using his evaluation harness on InfiAgent-DABench, Kevin found that two models running the same solver each answered about 87% of its 257 questions correctly. Qwen averaged 2.8 iterations per answer and MiniMax averaged 6.1.

Read the topics behind these talks

4 conference talks

Loading more talks and highlights…

References