Dr. Filippo Focacci: Building Optimization Systems That Survive Execution

Dr. Filippo Focacci is the co-founder and CEO of DecisionBrain, a software company that applies optimization and machine learning to planning and scheduling problems across workforce management, maintenance, supply chain, and logistics. He discovered operations research and AI at the same time as a student in the 1990s, wrote his master's thesis on combining constraint programming with OR, and built his career at ILOG before starting DecisionBrain in 2012.

His strongest argument is that optimization models are often credited with returns they have not yet produced. An objective function can estimate savings, but the estimate remains theoretical until the plan is carried out. When workers, machines, customers, or external conditions behave differently from the model’s assumptions, the expected return can disappear or turn negative. For Focacci, this makes execution data as important as the plan itself. A useful system must compare the proposed schedule with what is actually happening and revise decisions as conditions change.

The distinction is visible in real-time scheduling problems. London’s bike-sharing operation receives continuous information about bicycle movements and broken docking stations, requiring the schedule to be recalculated as supply shifts across the city. Airport cleaning presents a similar problem. An overnight schedule may assign a crew to clean a gate at a particular time, but a delayed flight can leave passengers occupying the area. Cleaning it then wastes labor. Once the passengers leave, the need may become more urgent than the original plan anticipated. The system has to interpret the new situation instead of continuing to follow a fixed sequence of tasks.

Focacci used the same approach in a Toyota inbound-logistics project. The first version replicated the rules and heuristics used by Toyota’s planners, producing a plan they could immediately recognize and assess. This built trust before more advanced optimization was introduced and gave the algorithm a strong starting solution. Reconstructing the manual process also exposed informal rules and accumulated judgment that a conventional requirements document would likely have missed.

This approach challenges the assumption that manual planning is simply an inefficient process waiting to be replaced. Experienced planners often hold knowledge that has never been formally documented. Reconstructing their decisions can reveal exceptions, priorities, and trade-offs that the initial mathematical formulation missed. The optimizer can then improve on an informed baseline instead of starting with an abstract version of the operation.

Focacci also resists choosing a method before understanding the problem. Constraint programming, mixed-integer programming, local search, and custom heuristics may all appear within the same system. In an early container-terminal project, constraint-programming heuristics generated possible loading and unloading sequences, while mixed-integer programming selected the combination most likely to reduce congestion. The choice of technique followed the structure of the problem.

His view of generative AI follows the same pattern. No single technique is sufficient for complex decision-making. Generative models can support parts of the process, but consequential decisions still require validation, explicit constraints, and human control. The focus is therefore on the complete decision system, where multiple methods work together, and each output is tested for safety, feasibility, and operational fit.

What is preventing your AI investments from delivering stronger business results?

TDWI Transform 2026 will run September 20-25 at the JW Marriott, Anaheim Resort, bringing together data, analytics, AI, and business leaders for six days of vendor-neutral training. The program includes 58 half- and full-day masterclasses, seven certificate bootcamps, hands-on labs, and a three-day Data & AI Leaders Summit. Sessions will cover AI and machine learning, BI and analytics, data governance, platform architecture, data strategy, and methods for deploying enterprise data and AI programs. Details →

ODSC AI West will take place October 27-29 at the Hyatt Regency San Francisco Airport, bringing together an expected 2,500 attendees, more than 250 speakers, and 300 hours of programming. Seven tracks will cover physical AI and robotics, agentic systems, AI engineering and evaluation, data infrastructure, machine learning, developer tools, and workflow automation, alongside hands-on training and an executive leadership summit. Confirmed speakers include Deepak Pathak of Skild AI, Pepijn Kooijmans of Hugging Face’s LeRobot project, Ganesh Balachandran of Waymo, and researchers and engineers from Anthropic, OpenAI, Google, Meta, Cursor, and LangChain. Details →

IBM surveyed 2,000 C-level technology executives across 33 geographies and 19 industries, in cooperation with Oxford Economics. Two-thirds of respondents are accountable for AI systems they do not fully control, 77% say adoption is already outpacing governance capabilities, and only 11% feel prepared for the 38% growth in deployed agents they expect by 2027. Organizations averaged 54 agent incidents last year requiring human correction, and 37% of high-severity cases resulted in data exposure. Organizations that embed control directly into their AI deployments see 25% fewer incidents, run 16 times more agents than those relying on manual governance, and deliver 18% higher operating margins. Read →

Benjamin Baer, Founding Member of the DecideWise community, argues that enterprises need a system of record for decisions. Knowledge management tools document what a policy is but lose why it was chosen, what trade-offs were accepted, and what circumstances justified exceptions. Baer draws a parallel to the legal system, where courts preserve dissenting opinions because the rejected arguments provide essential context when future courts revisit an issue. Enterprises have no equivalent, so when a strategy or business rule changes, the debate and the alternate paths considered are lost. His proposed mechanism pairs a decision trace, a record of how logic actually ran in a specific case, covering data inputs, exceptions granted, and sign-offs, with a decision graph that maps those traces across the organization. The value extends beyond audit. AI agents working from policy documents alone fail on edge cases, and they need historical traces showing where humans granted overrides and how conflicts between guidelines were actually settled. Baer closes on the commercial stakes. Systems of record are famously sticky once embedded, and the vendor that becomes the definitive record for enterprise decisions will hold a dominant position in corporate software. Read More →

Peter Cacioppi, Principal Operations Research Scientist at Decision Spot, joins Vijay Mehrotra and Michael Watson to explain how operations research teams can move optimization models from Jupyter notebooks into production software. He outlines a “tidy, tested, and safe” approach built around packaging Python code, exposing clear inputs and outputs, validating data before solving, and maintaining test datasets that confirm models still produce expected results after changes. He also discusses Ticdat, the open-source library he created to define and validate tabular inputs, and argues that LLMs are highly useful for Python development and documentation while mathematical models still require review by someone who understands the optimization logic.