Safety stock = z·√(L·σd² + d²·σL²), worked for one product: why lead-time spread dominates, why the two buffers do not add, and fill rate vs service level.
Badges
Badges come from GitHub once the author claims this feed.
Recent posts
every link opens the author’s post-
Safety stock when the supplier is late, worked with numbers
-
MAPE picked the forecast that RMSE called worse
MAPE formula worked on five days: 8.65% vs 7.14%, the reverse of RMSE.
-
Croston's method lost on the items it was built for
On 823 Walmart M5 items, Croston's method was 25 percent worse than a 28-day moving average on intermittent demand.
-
It won on the steady sellers and lost on the rest
An M5 case: a gradient-boosting forecast beats a moving average on smooth items and loses about 10 percent on intermittent and lumpy ones.
-
Economic order quantity and the reorder point, worked with numbers
Economic order quantity and the reorder point formula worked for one product: EOQ = √(2DS/H), why the cost curve is flat, safety stock, and when EOQ misleads.
-
Galois fields, worked by hand: GF(8) in one table
What a Galois field is, worked by hand: why mod 7 works and mod 8 fails, GF(8) built from x³ + x + 1, a full multiplication table, and GF(256) in QR and AES.
-
Forecast bias: the more accurate forecast was the biased one
Forecast bias worked through eight days of demand: mean error vs MAE and RMSE, percent bias, the tracking signal, and why bias is the cheapest error to fix.
-
Magnetic particle inspection, worked through with numbers
How magnetic particle inspection finds cracks in steel: flux leakage, why a crack must cut across the field, two magnetising directions, and what it cannot see.
-
Measurement uncertainty, worked through one micrometer reading
Measurement uncertainty worked through for a micrometer reading: type A and type B, the budget, root sum of squares, coverage factor k = 2, and what to fix.
-
The screwdriver did not fail. The handoff did.
Magneto timing on a Navion's piston engine, a screwdriver left in the spark plug hole, an engine locked solid, and the rules the hangar changed afterwards.
-
R squared and adjusted R squared, worked on eight points
R squared worked on eight points: what SS_res and SS_tot mean, why R² can go negative, how junk inputs raise it, and what adjusted R² corrects.
-
Surrogate models, worked through one cantilever beam
What a surrogate model is, worked on a cantilever beam: a quadratic fit vs a power law, error inside and outside the fitted range, and when to trust one.
-
Dye and fluorescent penetrant testing, worked through with numbers
How dye and fluorescent penetrant testing works: capillary pressure worked in numbers, the five steps, fluorescent vs visible dye, and what it misses.
-
Train/test split: one dataset, 45 splits, 45 different answers
What a train/test split measures, worked on ten points: every 80/20 split tried, test error from 0.12 to 3.19, leakage, and when to cross-validate instead.
-
Ultrasonic testing, worked through with numbers
How ultrasonic testing finds flaws: pulse-echo timing worked for a 10 mm aluminium plate, why cracks reflect, frequency, and what it cannot see.
-
Demand forecasting, worked through twelve weeks of numbers
Demand forecasting methods worked on twelve weeks: naive, moving average, exponential smoothing and Holt's trend, scored by MAE and bias, and what else matters.
-
Density altitude and true airspeed, worked in tables
Density altitude and true airspeed from the standard atmosphere: the formulas, a density altitude table by temperature, a TAS table, and the rule of thumb.
-
My validation ranked five models backwards
Bronze medal in a machine-learning competition on well geology: validation ranked five builds backwards (r = −0.745), and identical code scored 0.13 apart.
-
Eddy current testing, worked through with numbers
How eddy current testing finds cracks in aircraft metal: the coil, skin depth worked for aluminium, choosing the frequency, and what it cannot see.
-
Censored demand, worked through ten days of numbers
A worked example of censored demand: how stockouts bias the sales average, what you can still estimate exactly, and why high-margin products get stuck.
-
The newsvendor model, worked through with ten days of demand
Newsvendor formula: order up to where P(D ≤ q) reaches Cu/(Cu+Co).
-
Out-of-distribution failure, worked with five numbers
A worked example of out-of-distribution failure: a model that scores R² 0.83 on held-out data and −2.31 on new data, and why the real problem is subtler.
-
Group k-fold cross-validation, and the split that hides a failure
Why random k-fold leaks when rows share a group, and how GroupKFold fixes it.
-
How a pitot tube turns pressure into airspeed, in numbers
How a pitot tube measures airspeed: the Bernoulli equation behind it, a worked example in pascals and knots, indicated vs true airspeed, and blockages.
-
Reed–Solomon error correction, worked by hand
How Reed–Solomon error correction works, step by step: encoding as a polynomial, rebuilding lost symbols, fixing a wrong one, and why it needs a finite field.
-
RMSE vs MAE: two forecasts with the same average error
MAE counts every unit of error the same.
-
How we looked for cracks nobody can see
Non-destructive testing from the hangar: fluorescent penetrant that makes cracks glow, X-ray, and the tap test, and why aircraft parts are checked this way.
-
Five hours in an eye clinic, and why home monitoring is hard
Field research for a hackathon on remote AMD monitoring: the waiting room, the route in, five hours of one patient’s day, and why patients avoid apps.
-
Learning a Navion's cockpit, and what a pitot tube measures
An old Navion at an Air Force school: starting it from the cockpit, rudder, elevator and ailerons, and how a pitot tube turns air pressure into airspeed.
-
My load test had 1,000 players. None of them were late.
A load test that passed with 1,000 simulated players, the real room that broke anyway, and the decisions made in the ten minutes between.