will change into our digital assistants, serving to us navigate the complexities of the fashionable world. They are going to make our lives simpler and extra environment friendly.” Inspiring and fully 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 artistic. Arduous to disagree after I fortunately get pleasure from my morning espresso whereas Cursor is coding my unit checks. But, asking folks in my community how they use AI of their day-to-day, their solutions usually point out anecdotal use instances, anyplace from “I exploit it to inform bedtime tales to my son” (I assume that will not even be a use case in the event you 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 Knowledge Scientist, my thoughts goes forwards and backwards between two conclusions. The FOMO a part of me that doesn’t need to be late to the Robotic revolution social gathering, and the cynical one which thinks that there’s nonetheless a protracted method to go earlier than synthetic intelligence truly turns into clever. To search out out which aspect of my schizophrenic persona 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 because the starting of my profession and assessing how 2025 state-of-the-art AI fashions may have helped.
At this time, we return to 2018. I’m a candid summer season intern at one of the disruptive startups in America: Hire the Runway.
What the Venture was about
The Hire the Runway success heart in Secaucus, NJ, was once the largest dry cleansing facility in america.
Within the Summer time 2018, as an Operations Analyst intern, I used to be given a reasonably exhausting downside to consider: on a regular basis, the success heart was receiving hundreds of models again from throughout the nation. All of the gadgets needed to be first inspected, then would undergo a radical cleansing course of, earlier than being dried or receiving some particular therapies. This could possibly be:
- Recognizing if the garment was stained through the rental
- Urgent if it was too wrinkled and needed to be ironed
- Repairing if it had been broken
Most of those duties have been accomplished manually by completely different departments, and required specialised staff to be accessible as quickly as the primary batch of models have been reaching their division. Having the ability to predict days forward what quantity of models must be processed (and when) was essential for the success heart planning squad, in an effort to be sure that each operations staff 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 therapies, the place and when bottlenecks may seem, and understanding how the work accomplished at one division would affect the opposite departments.
The 2018 Resolution
At this level it’s possible you’ll surprise: given the complexity and the stakes of the venture, why was it within the fingers of a younger inexperienced intern? To be honest, throughout my 10-week summer season internship, I solely scratched the floor and wrote an insanely sophisticated Pyomo script that was later refined by a extra senior Knowledge Scientist, who spent two years on this venture alone.
However as you’ll be able to think about, the answer was this large optimization mannequin taking as an enter the inbound quantity forecast for on daily basis of the week, the typical 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 therapies. The principle constraints have 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 may 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 serious about seeing if with at the moment’s know-how, one other strategy could be attainable.
As a result of our strategy was, you could possibly say, fairly old fashioned, and required months and months of refinements and testing.
The principle restrict is the static side of the answer. If one thing surprising occurs through the week (e.g a snow storm that paralyzes the logistics in some components of the nation, delaying among the inbound quantity), a variety of assumptions of the mannequin need to be modified, and its outcomes have gotten out of date.
This can be a resolution that requires information scientists to go deep into the weeds, as an alternative of counting on an out-of-the-box framework, to depend on a variety of assumptions and to spend time sustaining and updating these assumptions.
Might AI provide you with a very completely different strategy? No.
For this explicit downside, 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 may suggest a framework with an AI agent appearing 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 resolve a posh optimization mannequin, and the sub-agents would want to speak because the state of affairs of 1 division can have an effect on all of the others.
Might AI considerably improve the “human-generated” resolution? 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 call out of it is likely to be exhausting for somebody with no sturdy understanding of linear programming. An LLM may assist interpret the outcomes and recommend concrete enterprise selections.
- Secondly, an LLM may assist react sooner to sure surprising conditions. It may for instance summarize info on occasions that would have an effect on the Operations, corresponding to unhealthy climate in some components of the nation or different points with suppliers, and as such, advocate when to rerun the planning mannequin. That’s assuming it has entry to good high quality information about these exterior occasions.
- Lastly, it’s attainable AI may have additionally helped with making actual time changes to the planning. As an example, it’s sometimes predictable based mostly on the garment traits whether or not they would require particular care (e.g a cotton shirt will all the time need to be ironed manually). Having a VLM scanning each garment on the receiving station may assist downstream departments perceive how a lot quantity they need to anticipate hours prematurely.
Might AI allow Knowledge Scientists to keep up 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 venture, though it may have enhanced the ultimate resolution. However it isn’t unimaginable to imagine that a couple of years (months?) from now, brokers with enhanced reasoning capabilities will likely be subtle sufficient to start out cracking most of these issues. Within the meantime, whereas AI may pace up mannequin iterations and changes, the human judgment on the core stays irreplaceable. This serves as a useful reminder that being a Knowledge Scientist isn’t nearly fixing mathematical or laptop science issues—it’s about designing sensible options that meet evolving, usually ambiguous and never so effectively outlined real-world constraints.
Article 100% human generated
