. Make better business decisions with data and AI—subscribe to The Decision Intelligence Lab Newsletter at https://decisionintelligencelab.substack.com/.
Optimization powers decisions in everything from healthcare to logistics, but for most practitioners it stays a "magical box": powerful, opaque, and locked behind PhD-level expertise. So what actually gets in the way of putting these models to work?
In this episode, hosts Vijay Mehrotra and Michael Watson sit down with Stanford's Postdoctoral Researcher Connor Lawless and Madeleine Udell (Assistant Professor at Management Science and Engineering Department, Stanford University) to unpack their paper "It Was a Magical Box": Understanding Practitioner Workflows and Needs in Optimization. Based on interviews with 15 optimization model developers, the research uncovers a surprising truth: the hardest part of operations research usually isn't the math. It's the people, the data, and the endless back-and-forth.
We dig into the six-stage workflow of building an optimization model, why nearly every project becomes an iterative "flywheel," why machine learning feels so much easier than OR, when "good enough" beats provably optimal solutions, and how LLMs might finally close the accessibility gap. Connor also shares what's next as he joins Percepta to work on the "last mile" of analytics.
Whether you build models for a living or just wonder why so many great models never make it into the real world, this conversation will change how you think about optimization in practice.
The Research Paper:
"It Was a Magical Box": Understanding Practitioner Workflows and Needs in Optimization (2025) - https://arxiv.org/abs/2509.16402
Timestamps
0:00 - Preview & Introduction
0:57 - Meet Connor Lawless and Madeleine Udell
2:04 - OR's accessibility problem vs. ML
3:07 - The six stages of building an optimization model
5:38 - Iteration as a flywheel: the "99% of the time" finding
7:45 - Why is ML so much easier than OR?
10:45 - The role of visualization and pattern recognition
12:00 - Optimization as a distribution-shift problem
13:24 - The big surprise: the human bottleneck, not the math
15:32 - Implications for teaching in the age of AI
18:00 - Handling uncertainty: data-scarce vs. data-rich problems
20:00 - Solver friction: Gurobi, CPLEX, and parameter tuning
22:20 - "Good enough" beats optimal
24:16 - Practitioner innovations: data and constraint validators
26:00 - Separating data from the model
27:37 - Preventing silent wrong answers
29:00 - Documentation, debugging, and LLMs as intermediaries
30:30 - Connor's next chapter: Percepta and the "last mile" of analytics
32:00 - If not us, then who?
Follow the show
Apple: https://podcasts.apple.com/in/podcast/the-decision-intelligence-lab/id1811085064
Spotify: https://open.spotify.com/show/0lFoAVKqJHTYSZNpeN61ou?si=0ae973aab0174b3b
Connect with guest
- Connor Lawless (Postdoctoral researcher, Stanford University): https://www.linkedin.com/in/connorlawless/
- Madeleine Udell (Assistant Professor, Management Science & Engineering, Stanford University): https://www.linkedin.com/in/madeleine-udell/
Connect with hosts
- Prof. Vijay Mehrotra (University of San Francisco): https://www.linkedin.com/in/vijay-mehrotra-ba9498/
- Prof. Michael Watson (Northwestern University): https://www.linkedin.com/in/michael-watson-07600a1
About the podcast
The Decision Intelligence Lab podcast delivers real-world insights for data professionals, business leaders, and anyone seeking to leverage data & AI for smarter decision-making & successful business outcomes.
For business inquiries, email at decisionintelligencepodcast@gmail.com
