will change into our digital assistants, serving to us navigate the complexities of the fashionable world. They may make our lives simpler and extra environment friendly.” Inspiring and utterly unbiased assertion from somebody who already invested billions on this new know-how.
The hype is actual for AI brokers, and billions are pouring in to construct fashions that may make us extra productive and extra inventive. Onerous to disagree after I fortunately get pleasure from my morning espresso whereas Cursor is coding my unit exams. But, asking individuals in my community how they use AI of their day-to-day, their solutions usually point out anecdotal use circumstances, anyplace from “I exploit it to inform bedtime tales to my son” (I assume that will not even be a use case for those who had extra creativeness) to “I exploit it to optimize my schedule” (Movement AI, please cease concentrating on me for the love of god).
As a Information Scientist, my thoughts goes backwards and forwards between two conclusions. The FOMO a part of me that doesn’t need to be late to the Robotic revolution get together, and the cynical one which thinks that there’s nonetheless a protracted option to go earlier than synthetic intelligence really turns into clever. To seek out out which facet of my schizophrenic character I ought to wager on, I’m going to make use of a easy but highly effective framework: reviewing all of the initiatives I’ve labored on for the reason that starting of my profession and assessing how 2025 state-of-the-art AI fashions might have helped.
At present, we return to 2018. I’m a candid summer time intern at some of the disruptive startups in America: Lease the Runway.
What the Mission was about
The Lease the Runway success middle in Secaucus, NJ, was once the largest dry cleansing facility in the US.
Within the Summer time 2018, as an Operations Analyst intern, I used to be given a fairly exhausting drawback to consider: on a regular basis, the success middle was receiving 1000’s of models again from throughout the nation. All of the objects needed to be first inspected, then would undergo an intensive cleansing course of, earlier than being dried or receiving some particular remedies. This may very well be:
- Recognizing if the garment was stained throughout the rental
- Urgent if it was too wrinkled and needed to be ironed
- Repairing if it had been broken
Most of those duties had been achieved manually by completely different departments, and required specialised staff to be obtainable as quickly as the primary batch of models had been reaching their division. With the ability to predict days forward what quantity of models must be processed (and when) was essential for the success middle planning squad, with the intention to guarantee that each operations group could be staffed appropriately.
The complexity of the circulation made it even trickier. It was not solely about predicting the inbound quantity, but in addition assessing what a part of this inbound quantity would require particular remedies, the place and when bottlenecks might seem, and understanding how the work achieved at one division would impression the opposite departments.
The 2018 Resolution
At this level you might surprise: given the complexity and the stakes of the challenge, why was it within the palms of a younger inexperienced intern? To be honest, throughout my 10-week summer time internship, I solely scratched the floor and wrote an insanely sophisticated Pyomo script that was later refined by a extra senior Information Scientist, who spent two years on this challenge alone.
However as you may think about, the answer was this large optimization mannequin taking as an enter the inbound quantity forecast for each day of the week, the common UPH (models per hour, i.e the variety of models that may be processed in an hour) at every division, and a few assumptions on the proportions of models that will require particular remedies. The primary constraints had been on the timing and regularity of the shifts, and the variety of full time contracts. The mannequin would then output an optimized labor planning for the week.
How AI might have helped
Let’s re-clarify issues first: you’ll not see phrases like “AI-enthusiast” or “LLM believer” in my LinkedIn bio. I’m fairly skeptical that AI will magically resolve all our issues, however I’m concerned with seeing if with as we speak’s know-how, one other strategy could be potential.
As a result of our strategy was, you may say, fairly old style, and required months and months of refinements and testing.
The primary restrict is the static side of the answer. If one thing surprising occurs throughout the week (e.g a snow storm that paralyzes the logistics in some elements of the nation, delaying a few of the inbound quantity), loads of assumptions of the mannequin should be modified, and its outcomes have gotten out of date.
It is a answer that requires knowledge scientists to go deep into the weeds, as a substitute of counting on an out-of-the-box framework, to depend on loads of assumptions and to spend time sustaining and updating these assumptions.
Might AI provide you with a totally completely different strategy? No.
For this specific drawback, you clearly want an optimization mannequin, and I’m but to examine an LLM with the ability to deal with a mannequin with such complexity. One might suggest a framework with an AI agent performing as a Basic Supervisor, and counting on sub-agents to deal with the planning of every division. However that framework would nonetheless require brokers to have instruments that enable them to unravel a fancy optimization mannequin, and the sub-agents would wish to speak because the scenario of 1 division can have an effect on all of the others.
Might AI considerably improve the “human-generated” answer? Attainable.
It’s at this level fairly apparent to me that LLMs wouldn’t make the issue trivial, however they might assist enhance the answer in a number of areas:
- To begin with, they might assist with reporting and choice making. The output of the optimization mannequin might need a enterprise sense, however making a choice out of it could be exhausting for somebody with no sturdy understanding of linear programming. An LLM might assist interpret the outcomes and counsel concrete enterprise selections.
- Secondly, an LLM might assist react sooner to sure surprising conditions. It might for instance summarize info on occasions that would have an effect on the Operations, comparable to unhealthy climate in some elements of the nation or different points with suppliers, and as such, suggest when to rerun the planning mannequin. That’s assuming it has entry to good high quality knowledge about these exterior occasions.
- Lastly, it’s potential AI might have additionally helped with making actual time changes to the planning. For example, it’s usually predictable primarily based on the garment traits whether or not they would require particular care (e.g a cotton shirt will all the time should be ironed manually). Having a VLM scanning each garment on the receiving station might assist downstream departments perceive how a lot quantity they need to anticipate hours prematurely.
Might AI allow Information Scientists to take care of and replace the mannequin? Sure!
It’s actually exhausting to disclaim that with instruments like Copilot or Cursor coding and sustaining this mannequin would have been simpler. I’d not have blindly requested Claude to code each constraint of the Linear Program from scratch, however with AI code editors being smarter than ever, modifying and testing particular constraints (and catching human errors!) could be simpler.
My conclusion is that an LLM in 2018 wouldn’t have trivialized the challenge, though it might have enhanced the ultimate answer. However it’s not unimaginable to imagine that a number of years (months?) from now, brokers with enhanced reasoning capabilities can be subtle sufficient to start out cracking a lot of these issues. Within the meantime, whereas AI might pace up mannequin iterations and changes, the human judgment on the core stays irreplaceable. This serves as a precious reminder that being a Information Scientist isn’t nearly fixing mathematical or pc science issues—it’s about designing sensible options that meet evolving, usually ambiguous and never so properly outlined real-world constraints.
Article 100% human generated
