What Changed in Data Science Job Market This Year

import pandas as pd

# Suppose an assistant wrote this to compute repeat-customer rate
def repeat_rate_ai(df):
    return (df["orders"] > 1).mean()          # assumes one row per customer

# Real data has one row per ORDER, so the result is wrong.
orders = pd.DataFrame({"customer": ["a", "a", "b", "c", "c", "c"],
                       "orders": [1, 1, 1, 1, 1, 1]})

def repeat_rate(df):
    per_customer = df.groupby("customer").size()
    return (per_customer > 1).mean()

print("AI version:", repeat_rate_ai(orders))       # 0.0 (wrong)
print("Checked version:", round(repeat_rate(orders), 2))  # 0.67

# A tiny evaluation harness for an LLM-based classifier.
# 'predict' stands in for any model call you make.
def predict(text: str) -> str:
    return "complaint" if "refund" in text.lower() or "late" in text.lower() else "other"

tests = [
    ("My refund has not arrived", "complaint"),
    ("Delivery was late again", "complaint"),
    ("Do you ship to Kochi?", "other"),
    ("The product broke in a week", "complaint"),
]
correct = sum(predict(t) == label for t, label in tests)
print(f"accuracy {correct}/{len(tests)}")
for t, label in tests:
    if predict(t) != label:
        print("MISS:", t)

from sklearn.datasets import make_classification
from sklearn.dummy import DummyClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score

X, y = make_classification(n_samples=1500, weights=[0.85], random_state=2)
for name, m in [("baseline", DummyClassifier(strategy="most_frequent")),
                ("logistic", LogisticRegression(max_iter=1000))]:
    f1 = cross_val_score(m, X, y, cv=5, scoring="f1").mean()
    print(f"{name:9} F1 = {f1:.3f}")
# Show both rows in your README; the baseline gives context.

# pip install streamlit scikit-learn joblib
# Run: streamlit run app.py
import joblib
import streamlit as st

st.title("Churn risk demo (synthetic data)")
model = joblib.load("model.joblib")

tenure = st.slider("Tenure (months)", 0, 60, 6)
spend = st.number_input("Monthly spend", 0, 10000, 500)
if st.button("Predict"):
    p = model.predict_proba([[tenure, spend]])[0][1]
    st.metric("Churn probability", f"{p:.0%}")

TITLES = ["Data Scientist", "Junior Data Scientist", "ML Engineer",
          "AI Engineer", "Analytics Engineer", "Applied Scientist",
          "Decision Scientist", "Data Analyst (ML)"]
FOR_FRESHERS = ["junior", "associate", "graduate", "trainee", "entry level"]

queries = [f'"{t}" {lvl}' for t in TITLES for lvl in FOR_FRESHERS[:2]]
print(len(queries), "queries, for example:")
for q in queries[:6]:
    print(" ", q)

import re

listing = """Responsibilities: build and evaluate ML models, write SQL,
deploy models as APIs, work with product teams, monitor model quality."""

signals = {
    "modelling": r"model|ml|predict",
    "sql": r"\bsql\b",
    "deployment": r"deploy|api|production",
    "llm": r"llm|genai|prompt",
}
for area, pat in signals.items():
    print(f"{area:11} {'yes' if re.search(pat, listing, re.I) else 'no'}")

Every two weeks, read 10 current listings you would apply to.
Record:
- Title and level
- SQL depth expected
- Deployment or MLOps mentioned?
- LLM / AI tools mentioned?
- Degree requirements

Adjust your learning to what appears most often.

Weeks 1-3   Strengthen SQL (CTEs, window functions)
Weeks 4-6   Project 1: model with baseline, error analysis, write-up
Weeks 7-8   Deploy Project 1 as an API or Streamlit demo
Weeks 9-10  Project 2: small LLM-based feature with an evaluation set
Weeks 11-12 Tailored applications across several titles; interviews

Prepare a two-minute story for each project:

1. Business question
2. Data and cleaning decisions
3. Baseline vs model, chosen metric
4. One error you found (including in AI-generated code)
5. How it was deployed or could be used
6. What you would improve next