than convincing somebody of a reality they can’t see in their very own knowledge.
Knowledge science and sustainability specialists face the identical downside: our ideas could also be too summary and theoretical, making them troublesome for decision-makers to narrate to.
I realized this the arduous manner whereas developping my startup!
After I revealed a case research on Inexperienced Stock Administration on TDS in 2024, I assumed the logic was stable and convincing, however the influence was restricted.
The article defined the mathematical idea behind it and used an precise case research to exhibit the sustainability advantages.
But it didn’t convert sceptics.
Buyer: “I’m certain it received’t work with our operations!”
Why? As a result of it wasn’t related to their knowledge and constraints.
So I made a decision to vary the strategy.
I packaged the simulation instrument in a FastAPI microservice and gave my prospects the flexibility to check the mannequin themselves utilizing an MCP Server related to Claude Desktop.
The target was to have them ask the LLM to run their very own eventualities, modify their parameters, and see how CO₂ emissions dropped in response to totally different stock insurance policies.
On this article, I’ll share the strategy I used for this experiment and the suggestions I acquired from a prospect, the Provide Chain Director of a retail group based mostly within the Asia Pacific area.
What’s Inexperienced Stock Administration?
On this part, I need to briefly clarify the idea of Inexperienced Stock Administration so you might have the context to grasp the instrument’s added worth.
Context: Stock Administration for a Retail Firm
Allow us to put ourselves in our Provide Chain Director’s footwear.
His groups (stock groups, warehouse and transportation operations) are liable for replenishing shops from a central distribution centre.
Once they want particular merchandise, shops robotically ship replenishment orders through their ERP to the Warehouse Administration System.
These automated orders comply with guidelines applied by the stock staff, referred to as the periodic evaluate “Order-Up-To-Stage (R, S)” coverage.
- The ERP is reviewing shops’ stock ranges, additionally referred to as stock available (IOH), each R days
- The delta between the goal stock S and the stock stage is calculated: Δ = S— IOH
- A Replenishment Order is created and transmitted to the warehouse with the amount: Q = S — IOH
After transmission, the order is ready on the warehouse and delivered to your retailer inside a selected lead time (LD) in days.
To be extra concrete, I share the instance above:
- R = 25 days: we evaluate the stock each 25 days as you’ll be able to see within the blue scatter plot
- S = 1,995 models: we ordered to succeed in this stage, as proven within the newest graph.
The stock groups within the methods often set these parameters, and the replenishment orders are robotically triggered.
What if we optimise these parameters?
Impacts on Logistics Operations
Based mostly on my expertise, these parameters are, more often than not, not set optimally..
The issue is that they considerably influence the effectivity of your warehouse and transportation operations.
This will increase carton and plastic consumption and reduces productiveness.
Within the instance above, objects are saved in cartons containing models that may be picked individually.
If the order amount is 5, the operator will:
- Open a field of 20 models and take 5 models ;
- Take a brand new field and put these things in it ;
- Palletise the packing containers utilizing plastic movie ;
The opposite influence is on truck filling price and CO2 emissions.
With a excessive supply frequency, you cut back the quantity per cargo.
This results in the usage of smaller vehicles that will not be full.
What can we do?
Aims of Inexperienced Stock Administration
We will take a look at a number of eventualities, with totally different operational parameters, to seek out the optimum setup.
For that, I’ve loaded buyer knowledge into the simulation mannequin
to check the instrument with actual eventualities.
Customers can modify a few of these parameters to simulate totally different eventualities.
class LaunchParamsGrinv(BaseModel):
n_day: int = 30 # Variety of days within the simulation
n_ref: int = 20 # Variety of SKUs within the simulation
pcs_carton: int = 15 # Variety of items per full carton
cartons_pal: int = 25 # Variety of cartons per pallet
pallet_truck: int = 10 # Variety of pallets per truck
ok: float = 3 # Security issue for security inventory
CSL: float = 0.95 # Cycle service stage goal
LD: float = 1 # Lead time for supply (days)
R: float = 2 # Assessment interval (days)
carton_weight: float = 0.3 # Carton materials weight (kg)
plastic_weight: float = 0.173 # Plastic movie weight per pallet (kg)These parameters embody:
n_dayandn_ref: outline the scope of simulationpcs_carton,cartons_pal,LDandpallet_truck: parameters linked to warehousing and transportation operationscarton_weight,plastic_weight: sustainability parametersR,okandCSL: parameters set by the stock staff
I need our Provide Chain Director to take a seat together with his groups (stock, warehouse, transportation and sustainability) to problem the established order.
If they should attain a selected goal, our director can:
- Problem his stock groups to seek out higher evaluate intervals (R), or cycle service stage (CSL) targets
- Ask the sustainability staff to seek out lighter carton supplies
- Redesign his warehouse operations to cut back the lead time (LD)
For that, we have to present them with a instrument to simulate the influence of particular adjustments.
That is what we’re going to do with the assist of an MCP Server related to Claude AI.
Demo of the Inexperienced Stock Administration AI Assistant
Now that we all know how this simulation instrument can add worth to my prospects, let me present you examples of analyses they’ve carried out.
These assessments have been carried out utilizing buyer knowledge over a simulation horizon of as much as 90 days.
I’ve replicated the questions and interactions utilizing anonymised dummy knowledge to keep away from sharing confidential info right here.
Onboarding of customers
I’ve related the MCP server to the Claude surroundings utilized by the Provide Chain managers to have them “play with the instrument”.
The bulk didn’t take the time to evaluate the preliminary case research and immediately requested Claude in regards to the instrument.
Hopefully, I’ve documented the MCP instruments to supply context to the agent, like within the toot launch_greeninv shared beneath.
@mcp.instrument()
def launch_greeninv(params: LaunchParamsGrinv):
"""
Launch an entire Inexperienced Stock Administration simulation.
This instrument sends the enter parameters to the FastAPI microservice
(through POST /grinv/launch_grinv) and returns detailed sustainability
and operational KPIs for the chosen replenishment rule (Assessment Interval R).
-------------------------------------------------------------------------
🌱 WHAT THIS TOOL DOES
-------------------------------------------------------------------------
It runs the total simulation described within the "Inexperienced Stock Administration"
case research, reproducing the habits of an actual retail replenishment system
utilizing a (R, S) Periodic Assessment Coverage.
The simulation estimates:
- Replenishment portions and order frequency
- Inventory ranges and stockouts
- Variety of full and blended cartons
- Variety of pallets and truck deliveries
- CO₂ emissions for every retailer and globally
- Carton materials and plastic utilization
- Operator productiveness (orderlines and items per line)
[REMAINDER OF DOC-STRING OMITTED FOR CONCISION]
"""
logging.data(f"[GreenInv] Working simulation with params: {params.dict()}")
attempt:
with httpx.Consumer(timeout=120) as consumer:
response = consumer.publish(LAUNCH, json=params.dict())
response.raise_for_status()
consequence = response.json()
last_run = consequence
return {
"standing": "success",
"message": "Simulation accomplished",
"outcomes": consequence
}
besides Exception as e:
logging.error(f"[GreenInv] Error throughout API name: {e}")
return {
"standing": "error",
"message": str(e)
}I used to be fairly happy with Claude’s introduction to the instrument.
It begins with the introduction of the core capabilities of the instruments from an operational standpoint.
Shortly, our director began to ship me lengthy emails with questions on the best way to use the instrument:
- arrange the parameters?
- Who ought to I contain on this train?
My preliminary reflex was to reply: “Why don’t you ask Claude?”.
That is what they did, and the outcomes are wonderful. Claude proposed a framework of research.
This framework is sort of excellent; I’d simply have put the lead time (LD) additionally within the scope of the Warehouse Supervisor.
Nonetheless, I must admit that I’d by no means have been in a position to generate such a concise and well-formatted framework by myself.
Then, Claude proposed a plan for this research with a number of phases.
Let me take you thru the totally different phases from the person’s perspective.
Part 1: Baseline Evaluation
I suggested the staff to repeatedly ask Claude for a pleasant dashboard with a concise govt abstract.
That’s what they did for Part 1.
As you’ll be able to see within the screenshot above, Claude used the MCP Server instrument launch_greeninv to run an evaluation with the default parameters outlined within the Pydantic mannequin.
With the outputs, it generated the Government Abstract for our director.
The abstract is concise and straight to the purpose.
It compares the outputs (key efficiency indicators) to the targets shared within the MCP docstring and the grasp immediate.
What in regards to the managers?
Then it generated team-specific outputs, together with tables and feedback that clearly highlighted essentially the most vital points, as proven within the instance beneath.
What’s attention-grabbing right here is that our warehouse supervisor solely talked about the goal items per line in a earlier message.
Meaning we will have the instrument study not solely from the MCP’s instruments docstrings, grasp immediate, and Pydantic fashions, but additionally from person interactions.
Lastly, the instrument demonstrated its means to have a strategic strategy, offering mid-term projections and alerting on the important thing indicators.
Nonetheless, nothing is ideal.
When you might have weak prompting, Claude by no means loses the chance to hallucinate and suggest selections outdoors the scope of the research.
Allow us to proceed the train, following Anthropic’s mannequin, and proceed to Part 2.
Part 2: Situation Planning
After brainstorming with its staff, our director collected a number of eventualities from every supervisor.
What we will see right here is that every supervisor needed to problem the parameters centered on their scope of accountability.
This thought course of is then transcribed into actions.
Claude determined to run the six eventualities listed above.
The problem right here is to compile all the outcomes into an artificial, insight-driven abstract.
Within the case research revealed in 2024, I centered solely on the primary three eventualities, analyzing every efficiency indicator individually.
What about Claude?
Claude was smarter.
Though we had the identical sort of information available, it produced one thing extra “cross-functional” and decision-driven.
- We now have business-friendly names for every state of affairs which are comprehensible throughout features.
- Every state of affairs is linked to the staff that pushed for it.
Lastly, it offered an optimum state of affairs that may be a consensus between the groups.
We’re even supplied with a scorecard that explains to every staff why the state of affairs is greatest for everyone.
For a extra detailed breakdown of the agent’s outputs, be happy to take a look at this tutorial:
Conclusion
A brand new hope for the idea of Inexperienced Stock Administration
After a few weeks of experimentation, the Provide Chain Director is satisfied of the necessity to implement Inexperienced Stock Administration.
The one bottleneck right here is on their aspect now.
With Claude’s assist, our 4 managers concerned within the research understood the influence of their roles on the distribution chain’s total effectivity.
This helps us at LogiGreen onboard Provide Chain departments for complicated optimisation workouts like this one.
For my part, it’s simpler to conduct a inexperienced transformation when all groups have possession and sponsorship.
And the one method to get that’s to verify all people understands what we’re doing.
Based mostly on the preliminary outcomes of this modest experiment, I feel we’ve got discovered a wonderful instrument for that.
Would you like different case research utilizing MCP Server for Provide Chain Optimisation?
AI Agent for Provide Chain Community Optimisation
In one other article revealed on In direction of Knowledge Science, I share the same experiment centered on the Provide Chain Community Design train.
The target right here is extra macro-level.
We need to decide where items are produced to serve markets on the lowest value in an environmentally pleasant manner.
Whereas the algorithm differs, the strategy stays the identical.
We attempt a number of eventualities with parameters that favour totally different groups (finance, sustainability, logistics, manufacturing) to succeed in a consensus.
Like right here, Claude does a terrific job in synthesising the outcomes and offering data-driven suggestions.
For extra particulars, you’ll be able to watch this video.
About Me
Let’s join on Linkedin and Twitter. I’m a Provide Chain Engineer who makes use of knowledge analytics to enhance logistics operations and cut back prices.
For consulting or recommendation on analytics and sustainable provide chain transformation, be happy to contact me through Logigreen Consulting.
In case you are fascinated with Knowledge Analytics and Provide Chain, have a look at my web site.
