How AI Is Used in Indian Logistics Companies Today

import pandas as pd

# hub_volumes.csv: date, hub, parcels
df = pd.read_csv("hub_volumes.csv", parse_dates=["date"])
df = df.sort_values(["hub", "date"])

df["dow"] = df["date"].dt.dayofweek
df["month"] = df["date"].dt.month
df["lag_7"] = df.groupby("hub")["parcels"].shift(7)
df["rolling_28"] = (df.groupby("hub")["parcels"]
                      .transform(lambda s: s.shift(1).rolling(28).mean()))

# A calendar of known sale days helps capture spikes
sale_days = pd.to_datetime(["2026-10-10", "2026-10-11"])
df["is_sale"] = df["date"].isin(sale_days).astype(int)
print(df.dropna().head())

from sklearn.ensemble import HistGradientBoostingRegressor
from sklearn.metrics import mean_absolute_error

# continuing from the features tab
data = df.dropna()
features = ["dow", "month", "lag_7", "rolling_28", "is_sale"]
train = data[data["date"] < "2026-09-01"]
test = data[data["date"] >= "2026-09-01"]

model = HistGradientBoostingRegressor(random_state=0)
model.fit(train[features], train["parcels"])
pred = model.predict(test[features])
print("MAE:", round(mean_absolute_error(test["parcels"], pred), 1))

# Compare with the simple baseline: last week's value (lag_7)
print("Baseline MAE:", round(mean_absolute_error(test["parcels"], test["lag_7"]), 1))

import math

# A simple greedy route: always visit the nearest unvisited stop.
# Real systems use stronger solvers, but this shows the idea.
stops = {"Hub": (0, 0), "A": (2, 3), "B": (5, 1), "C": (1, 6), "D": (6, 5)}

def dist(p, q):
    return math.dist(stops[p], stops[q])

route, current = ["Hub"], "Hub"
remaining = set(stops) - {"Hub"}
while remaining:
    nxt = min(remaining, key=lambda s: dist(current, s))
    route.append(nxt)
    remaining.remove(nxt)
    current = nxt

total = sum(dist(a, b) for a, b in zip(route, route[1:]))
print(route, round(total, 2))

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_absolute_error

# trips.csv: distance_km, hour, stops, area_type, minutes
trips = pd.read_csv("trips.csv")
X = pd.get_dummies(trips[["distance_km", "hour", "stops", "area_type"]])
y = trips["minutes"]

X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.2, random_state=0)
model = RandomForestRegressor(n_estimators=200, random_state=0).fit(X_tr, y_tr)
print("ETA MAE (minutes):", round(mean_absolute_error(y_te, model.predict(X_te)), 1))

import re

raw = "Flat 12B, near Shiv Mandir, sec-62 noida up 201301"

def extract(address):
    pin = re.search(r"\b[1-9]\d{5}\b", address)
    sector = re.search(r"\bsec(?:tor)?[\s-]*(\d+)", address, re.I)
    landmark = re.search(r"near ([^,]+)", address, re.I)
    return {
        "pincode": pin.group() if pin else None,
        "sector": sector.group(1) if sector else None,
        "landmark": landmark.group(1).strip() if landmark else None,
    }

print(extract(raw))
# {'pincode': '201301', 'sector': '62', 'landmark': 'Shiv Mandir'}

from difflib import get_close_matches

known_localities = ["Indirapuram", "Vaishali", "Raj Nagar", "Crossings Republik"]

typed = ["indrapuram", "vaishalli", "rajnagar extn", "crosing republic"]
for t in typed:
    match = get_close_matches(t.title(), known_localities, n=1, cutoff=0.6)
    print(f"{t:18s} -> {match[0] if match else 'NEEDS REVIEW'}")

# Production systems use trained models, but low-confidence
# matches should still go to a human for review.

AI use case canvas

Problem: ____________________________
Who is affected (hub, rider, customer): __________
Current approach and its cost: __________________
Data available: ______________________________
AI method (forecast / classify / optimise / NLP / vision):
Success metric: ______________________________
Fallback when AI is unsure: ___________________
How results reach the decision maker: __________

Portfolio project ideas (use public or synthetic data)

1. Daily parcel volume forecast with a sale-day feature
2. ETA prediction from trip features with error analysis
3. Address parser that extracts pin code and landmark
4. Simple route optimiser for 10-20 stops
5. Delay-risk classifier with precision/recall trade-off

Always compare with a simple baseline and explain
what decision the model would support.

Before deploying logistics AI

[ ] Baseline comparison done
[ ] Tested on recent, unseen periods
[ ] Behaviour checked on sale days and holidays
[ ] Low-confidence cases sent to humans
[ ] Personal data minimised and protected
[ ] Monitoring for drift after launch
[ ] Operations team trained on how to override