Kalpa Logistics

Kalpa Logistics

Last-mile delivery and warehousing for every other unit. A cost centre that everybody blames, measured on a promise it does not fully control.

Code LOG. The home of time, capacity and the tail: censored outcomes, quantiles rather than means, and the long right-hand end of a distribution that nobody looks at.

This page is ahead of the lock

Client zero v2.2 schedules Build 1 in Health, Build 2 in Financial Services and Build 3 in Connect, and names Logistics only as a unit a later build visits. Weeks 10 to 20 are outline until they are detailed on go-ahead, so nothing below is scheduled to a week yet. The unit line, the KPIs and the situations here are usable as transfer examples and as Situation Bank material; they are not a build-week brief, and no day pack should treat them as one.

---

1. Who runs it, and what each of them wants

Role	Measured on	What they want to hear	What they will resist
Head of Last Mile	On-time percentage and cost per delivery	That the network is improving	A number that keeps undelivered orders in the denominator
Warehouse Operations Manager	Picks per hour, dispatch cut-off adherence	That the hub is efficient	A per-line comparison against a hub that handles a different mix
Head of Network Planning	Capacity against forecast	That the forecast was right	A forecast measured per hub per day rather than nationally per month
Head of Retail (as internal customer)	The delivery promise on the product page	An aggressive number, because it converts	A conservative quantile
Head of Customer Experience	Complaints about delivery	A number that predicts complaints	Being told the mean is not the thing that generates them

The tension that generates most cards here: the promise shown to a customer is a commercial decision with a cost on both sides, and it is made by a unit that does not carry the cost of missing it. Logistics is where a learner first meets a question with two defensible answers.

---

2. The data estate

What exists	The catch it carries
Order, dispatch and delivery timestamps	About four percent of rows have no delivery timestamp, and those are not missing data
Pincode, courier partner, weight band, status	Pincodes moved to a partner stop appearing in the dataset entirely
Exception codes on failed attempts	The code is chosen by the driver, and "customer unavailable" is the cheapest one to select
Driver location pings	High frequency, which tempts a system into replanning more often than a human can follow
Warehouse pick and pack times per line	Hubs handle different mixes, so per-hub averages compare different work

The absence that defines the unit: nothing records whether a failed attempt was the customer's fault or the driver's. The single most consequential field in last-mile operations is self-reported by the person being measured.

---

3. The KPIs, and how each one is gamed

KPI	What people say it is	The denominator that decides it	How it breaks or gets gamed
On-time delivery	Delivered by the promised date	Orders delivered, which quietly excludes orders never delivered	The worst outcomes leave the metric entirely, so the number rises as failures rise
Average delivery time	Mean hours from order to doorstep	Delivered orders only, and a mean rather than a quantile	A promise beaten half the time is broken half the time. The business needs a 90th percentile.
First attempt success	Delivered on the first visit	Attempts, as recorded by the driver	The exception code is the driver's choice
Cost per delivery	Total cost divided by deliveries	Per parcel or per order, and multi-parcel orders differ	Splitting an order into parcels improves the per-unit number and costs more in total
Capacity utilisation	Volume against capacity	Per hub, per shift	A national monthly average hides the two shifts a week that fail
SLA by pincode	Service level in a region	The pincodes still in the network	Move the slowest pincodes to a partner and the average improves with no operational change
Exception rate	Deliveries that went wrong	What counts as an exception	Reclassification is free and instant

The KPI this unit exists to teach: a quantile with a stated confidence, and a separate estimate of the probability of never arriving. Those two together are a promise. A mean is a wish.

---

4. Case families, mapped to modules

Module	Case family here	The teaching version
1 Foundations of AI and Data	On-time rates, delivery time distributions, the shape of a right tail	The first time a median and a mean tell different stories about the same file
2 Applied Machine Learning	Delivery-time prediction with censoring, demand and capacity forecasting, slot optimisation	Quantile loss rather than squared error, and censored rows handled rather than dropped
3 Deep Learning and Neural Networks	Sequence models over route and scan events	Where the extra capacity has to earn its place against a strong baseline
4 Natural Language Processing	Address parsing and normalisation for Indian addresses	Accuracy measured on the addresses a human was employed to fix, not the easy ones
5 and 6 Generative AI	Summarising a day's exceptions for a hub manager	An output whose only reader has ninety seconds
7 Production AI	Latency budgets for a driver app, and a p99 that a median hides	Where a service level stops being a model metric
8 and 9 Agentic AI	Exception-handling agents, replanning, and the constraint that a human needs a stable plan	Autonomy limited by human tolerance rather than by capability

---

5. The situations hosted here

Page	Cards
Analytics	S-LOG-AN-13 to S-LOG-AN-15
Machine learning	S-LOG-ML-03 (the orders that have not arrived yet), S-LOG-ML-13
GenAI and agents	S-LOG-GA-11 to S-LOG-GA-13

---

6. What Logistics must never be used for

1. A first lesson in modelling. Censoring and quantile loss are hard, and a room meeting its first model should not meet them on the same day.
2. A route optimisation that becomes an operations research course. The teaching point is the constraint nobody wrote down, and the algorithm is not the subject.
3. Blaming a driver. The exception-code card is about a measurement system that asks the measured party to report on itself. That is a systems point, and it stays one.