01 // safety_stock.py
Safety stock & reorder points
Sizes safety stock and ROP under both P1 cycle-service and P2 fill-rate definitions — and stops to ask which one the business actually means before calculating.
01.2 // SUPPLY CHAIN · LIVE PILOT
DOMAIN KNOWLEDGE,
MADE CALLABLE.
Replen is a supply-chain planning skill built from two materials: planning theory, and the judgment that only appears after you have worked the exceptions. It calculates with your data, states every assumption, and leaves the decision with the human.
01
KNOWLEDGE PACK
05
PYTHON CALCULATORS
02
CSV TEMPLATES
00
ERP WRITE-BACKS
// why I built it
A planner is more than a formula library.
A general AI can explain EOQ. That is not the hard part. The hard part is knowing when EOQ is the wrong question, which service-level definition the business is using, whether the variability belongs to demand or forecast error, and when the data is too weak for a confident answer.
Replen is my attempt to write that layer down. It turns tacit planning judgment into a repeatable workflow: classify the question, expose missing inputs, calculate through tested tools, show sensitivity, name the trade-off, and say where the recommendation stops being trustworthy.
THE OPERATING RULE
Numbers without assumptions are forbidden.
Every answer carries its formula, inputs, missing-data flags, applicability boundary and the trade-off being purchased.
// what it can do
Five calculations.
One decision system.
The calculators are not five isolated demos. Replen uses them inside a planning conversation — first checking whether the problem and data fit the method, then turning the result into an operating recommendation.
02 // forecast_metrics.py
Forecast diagnostics
Measures MAD, RMSE, MAPE, WMAPE, bias and tracking signal, then checks whether the SKU is smooth, erratic, intermittent or lumpy before recommending which variability measure belongs in the policy.
03 // croston.py
Intermittent demand
Uses Croston and the bias-corrected SBA method for demand with too many zero periods — instead of forcing every slow mover through a normal-distribution formula.
04 // eoq.py
EOQ with reality attached
Calculates the theoretical EOQ, then compares MOQ, pack and candidate quantities so the planner can see the real cost penalty of rounding rather than treating one formula as an order instruction.
05 // abc_xyz.py
ABC–XYZ policy design
Segments items by value and variability, then suggests a policy by class — where precision is worth the effort, where min–max is enough, and where not stocking may be the honest answer.
// how I made it
From experience
to a reusable skill.
- 01
Start with repeated decisions
I began with the questions a planner meets again and again: how much safety stock, why the stockout keeps returning, whether the forecast is biased, and whether an MOQ is actually expensive.
- 02
Write down the judgment
The useful part was not another glossary. I encoded the decision routes: calculate, diagnose, design, explain a concept, or translate planning logic into system parameters — each with the questions that must be answered first.
- 03
Move the maths into tools
Five Python calculators became the numerical spine. They are dependency-free, accept single-item or CSV inputs where appropriate, and keep arithmetic separate from the language model’s prose.
- 04
Add the traps textbooks hide
Service level can mean P1 or P2. Forecast-error RMSE is not raw demand standard deviation. Intermittent demand breaks normal safety-stock logic. These checks became gates, not footnotes.
- 05
Separate public knowledge from private context
Reusable planning knowledge, original examples and generic templates form the public pilot. A red-line check keeps employer, customer and real-person identifiers out before anything can be shared.
THE ARCHITECTURE
One source.
Several useful forms.
- Instructions — the role, answer protocol, checks and voice.
- Knowledge — the inventory and replenishment pilot pack, organised around planner jobs rather than textbook chapters.
- Tools — five Python calculators for reproducible arithmetic.
- Templates — two CSV schemas that make missing inputs visible before the model invents them.
THE BOUNDARY
Decision support,
not autopilot.
- It does not write orders or parameters back to an ERP.
- It does not invent missing lead times, forecasts or service targets.
- It does not turn a textbook formula into a business guarantee.
- It does label theory, operator judgment and data-dependent uncertainty separately.
INDEPENDENT EDUCATIONAL PROJECT · VALIDATE AGAINST YOUR OWN DATA BEFORE ACTING
THE SIGNAL // DOMAIN × AI × GOVERNANCE
The real build is the judgment layer.
Replen is evidence that domain experience can be turned into a system: explicit rules, reusable tools, structured inputs, privacy gates and a human decision boundary — not just a longer prompt.
→ read the field note: how I graded and rebuilt the skill