Will be our digital assistant to help us navigate the complexity of the modern world. They will make our lives easier and more efficient. “An inspiring, completely unbiased statement from someone who has invested billions of dollars in new technology.
For AI agents, the hype is real, and billions of dollars are building models, which will make us more productive and creative. It’s hard to disagree with morning coffee when I test it in my unit with cursor code. But, asking people on my network how to use AI in everyday use of AI, their answers often refer to anecdotal use cases, from anywhere “I use it to tell my son’s bedtime story” (I guess if you have more imagination, that’s not even use cases) to “I use it to optimize my schedule” (Motion AI, please stop targeting me for love for God).
As a data scientist, my mind walks back and forth between two conclusions. My FOMO part doesn’t want to be late for the Robotic Revolution, and cynical people think there’s a long way to go before AI can really get smart. To find out which side of the schizophrenia personality I should bet on, I will use a simple and powerful framework: look back on all the projects I have worked on since the beginning of my career and evaluate how the 2025 state-of-the-art AI models can help.
Today, we go back to 2018. I’m a candid summer intern at one of the most destructive startups in the United States: renting a runway.
What is this project
The runway fulfillment center rental in Secaucus, New Jersey was once the largest dry cleaning facility in the United States.
In the summer of 2018, as an operations analyst intern, I had a very difficult problem: every day, the fulfillment center received thousands of units from all over the country. All items must be checked first and then a thorough cleaning process is performed before drying or receiving some special treatment. This might be:
- Discover whether the clothing is dyed during rental period
- If it is too wrinkled, it must be ironed
- Is the repair damaged?
Most of these tasks are done manually by different departments, requiring professional workers to be provided immediately after the first units arrive at their departments. Being able to predict the number (and when) of units that must be processed in the coming days is critical to fulfilling the Center Planning Squad to ensure that every operational team is appropriate.
The complexity of flow makes it even more tricky. This is not only about predicting inbound volumes, but also about evaluating which portion of that inbound volume requires special treatment, when and where bottlenecks appear, and understanding how one department’s work will affect others.
2018 Solutions
At this point, you might be wondering: Given the complexity and bets of the project, why is this in the hands of a young, inexperienced intern? To be fair, during my 10-week summer internship, I scratched the surface and wrote a crazy complex Pyomo script that was later refined by a more advanced data scientist who spent only two years on this project.
But, as you can imagine, the solution is to take this huge optimization model as input, as an inbound forecast for every day of the week, average UPH (units per hour, i.e. the number of units that each department can process in an hour), and some assumptions about units that require a specific processing. The main limitations are the time and regularity of shifts and the number of full-time contracts. The model will then output an optimized labor plan this week.
How AI can help
Let’s repost things first: You won’t see words like “Ai-therusiast” or “LLM Believer” in my LinkedIn Bio. I’m very suspicious that AI will magically solve all of our problems, but I’m interested in knowing if today’s technology is possible.
You can say that because our approach is quite old school and requires months of refinement and testing.
The main limit is the static aspect of the solution. If something unexpected happens within a week (for example, a blizzard that paralyzes logistics in some parts of the country, delaying some inbound volume), many of the assumptions of the model must be changed, and the result becomes obsolete.
This is a solution that requires data scientists to go deep into the weed instead of relying on out-of-the-box frameworks to rely on many assumptions and take the time to maintain and update those assumptions.
Can AI come up with a completely different approach? No.
For this particular question, you obviously need an optimization model, and I haven’t read about the ability of LLM to handle the model with such complexity. People can use AI agents as the framework for general managers and rely on child agents to handle plans for each department. However, the framework still requires agents to have tools to enable them to solve complex optimization models, and child agents need to communicate because the situation in one department affects all others.
Can AI significantly enhance the solution of “human generation”? Possible.
In my opinion, LLMS doesn’t make the problem trivial, but they can help improve solutions in multiple areas:
- First, they can help with reporting and decision making. Optimizing the output of a model can be business-oriented, but making a decision can be difficult for people who don’t understand linear programming. LLM can help explain the results and suggest specific business decisions.
- Second, LLM can help respond faster to certain unexpected situations. For example, it can summarize information about events that may affect operations, such as bad weather in certain parts of the country or other issues with suppliers, so it is recommended when to rerun the planning model. This is assuming it has access to high-quality data about these external events.
- Finally, AI may also help make real-time adjustments to plans. For example, it is often possible to predict whether special care is required based on the characteristics of the garment (for example, a cotton shirt must always be ironed manually). VLM scans of each garment at the receiving station can help the downstream departments understand the amount they should expect in advance.
Can AI enable data scientists to maintain and update models? Yes!
It’s really hard to deny that it would be easier to use a copy or cursor encoding and maintaining a tool like this model. I won’t blindly ask Claude to encode every constraint of a linear program from scratch, but as AI code editors are smarter than ever, modifying and testing specific constraints (and catching human errors!) will become easier.
My conclusion is that while the project can enhance the final solution, LLM in 2018 won’t make the project trivial. However, it is not impossible to believe that from now on the years (months?), an agent with enhanced reasoning will be complex enough to start cracking these types of problems. At the same time, although AI can speed up the iteration and adjustment of models, the core human judgment is still irreplaceable. It is a valuable reminder that becoming a data scientist is not just about solving mathematical or computer science problems, but also about designing practical solutions that are in line with evolving and evolving, often ambiguous and less well-defined real-world constraints.
100% human-generated terms
