How I Spot a Copy-Pasted ML Engineer Project in Minutes

# Count commits and see how the work grew
git log --oneline | wc -l
git log --stat --reverse | head -60

# Who wrote what, and when
git shortlog -sn
git log --format="%ad %s" --date=short | head -20

# What changed in the notebook over time
git log --follow -p -- notebooks/model.ipynb | head -80

a1c3e9f  add raw data loader and basic EDA
4b7d201  baseline: majority class and logistic regression
9e02aa1  fix: target leaked via 'refund_date', drop column
c55f310  add time-based split, scores drop but are honest
0d81b2e  try gradient boosting, small gain on recall
7fa4c09  error analysis by city, add README findings

Even without opening a file, this history tells a story
of a person who found a bug and fixed it.

Interviewer: Why did you use StandardScaler before the
random forest?

Candidate: Because scaling improves the model.

Interviewer: Does a random forest need scaling?

Candidate: I think... yes, all models need it.

(Tree models split on thresholds, so scaling does not
change them. The step was copied from a template that
also used logistic regression.)

Interviewer: Why did you use StandardScaler before the
random forest?

Candidate: Honestly, the forest does not need it, since
trees split on thresholds. I kept the scaler in the pipeline
because I compared the forest with logistic regression,
which does need scaled inputs, and I wanted one pipeline
for both. I checked that the forest's score did not change
with or without it.

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

data = load_breast_cancer(as_frame=True)
X, y = data.data, data.target
groups = {
    "mean_*": [c for c in X.columns if c.startswith("mean")],
    "worst_*": [c for c in X.columns if c.startswith("worst")],
    "*_error": [c for c in X.columns if c.endswith("error")],
}

def score(cols):
    model = make_pipeline(StandardScaler(), LogisticRegression(max_iter=2000))
    return cross_val_score(model, X[cols], y, cv=5, scoring="recall").mean()

all_cols = list(X.columns)
print("all features: %.3f" % score(all_cols))
for name, cols in groups.items():
    kept = [c for c in all_cols if c not in cols]
    print("without %-8s %.3f" % (name, score(kept)))

# pip install fastapi uvicorn scikit-learn joblib
# Run: uvicorn serve:app --reload
import joblib
import numpy as np
from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI()
model = joblib.load("model.joblib")  # a fitted sklearn pipeline

class Features(BaseModel):
    values: list[float]

@app.post("/predict")
def predict(f: Features):
    x = np.array(f.values).reshape(1, -1)
    prob = float(model.predict_proba(x)[0, 1])
    return {"probability": round(prob, 4)}

PROJECT: ____________________

[ ] I can say why I chose this dataset
[ ] I changed at least one thing the tutorial did not do
[ ] I have a baseline score written down
[ ] I can explain every preprocessing step in one line
[ ] I ran at least one ablation
[ ] I can name two things that did not work
[ ] My README has my own numbers and limitations
[ ] My commit history shows real steps

Score under 6? Spend a weekend on this project
before you list it again.

STEP                       WHY I DID IT              EVIDENCE
-------------------------  ------------------------  -----------------
Dropped column ________    ________________________  score before/after
Filled missing with _____  ________________________  ______________
Encoding: ______________   ________________________  ______________
Model: _________________   ________________________  vs baseline ____
Metric: ________________   ________________________  business reason
Split: _________________   ________________________  matches real use?

"This project started from a public tutorial on ____.
I changed it by ____ because I wanted to ____.

My baseline was ____ with a score of ____.
My final model scored ____.

The step that mattered most was ____: removing it dropped
the score to ____. A step that did not help was ____.

What failed: ____. What I would do next: ____."