STOP 01
Farm gate · interior Sindh
Problem type — Measurement
The number that doesn't exist
Start with the headline everyone quotes: a third of the crop is lost between farm and buyer.2 Then ask the uncomfortable question — how does anyone know it is 30%? Nobody weighed it. There is no scanner at the farm gate, no timestamp, no record. The figure is an estimate laid over an informal economy where the real data lives in a broker's memory and a paper challan.
So the first hard problem is not analysis at all — it is measurement. In a formal economy the variable you would slice on already exists inside an ERP. Here it was never captured. That is the layer where Pakistani supply-chain companies actually get built: not better analytics on clean data, but the capture that makes data exist. Trukkr digitising trucking,3 Haball digitising B2B invoicing3 — measurement plays before they are anything else.
Farm-gate readout
Post-harvest loss30%+
Sector share of GDP~20%
Sensors at the farm gate0 — the variable was never recorded
STOP 02
The aggregator · the mandi
Problem type — Simpson's paradox7
The aggregate that lies
The crate reaches an aggregator buying from two supplier networks. On paper, network A passes quality at 90%, network B at 75%. The obvious move is to drop B. Make the room commit to it — out loud.
Now split by the one variable nobody logged: cold-chain availability on the route. B beats A on cold-chain routes and on routes with no cold chain. B only looked worse because it serves the hard rural terrain where no cold chain exists. The average measured B's terrain, not B's performance — and the buyer who consolidates to A just fired the better operator. At national scale, that single misjudgment is how smallholders get pushed out of formal supply.
Aggregate — the lie
Network A90%
Network B75%
→
Split by cold chain — the truth
Cold-chain routesB > A
No-cold-chain routesB > A
STOP 03
On the road · N-5 to Karachi
Problem type — Variance & tail risk
The average that hides the risk
Now the crate goes on a truck, and the question shifts from quality to time. Two carriers: one averages 10 days, one averages 14. The 10-day carrier looks better — until you show the spread. The fast one is 10 days give or take 8; the slow one is 14 give or take 1.
For a perishable crate with an export ship to catch, the reliable-but-slower carrier is worth more: you can plan around it, and the fast one strands your fruit more than a quarter of the time. Variance beats the mean. And with perishables it compounds — value does not just wait, it decays with every extra day, so the tail of that distribution is not a delay, it is a total loss. The number people optimise is the wrong number; the number that matters is not on the dashboard.
Carrier A — looks faster10 days ±8
Carrier B — actually plannable14 days ±1
0 days
day 16 · ship cutoff
24
average transit
typical spread
past the ship — total loss
STOP 04
Exporter → retailer abroad
Problem type — The bullwhip effect8
The signal that distorts as it climbs
Zoom out to the whole chain. A small bump in real consumer demand becomes a larger order from the retailer, a larger one from the exporter, and a panic order back at the farm — each link padding against uncertainty in the one above it. This is the bullwhip effect,8 and it is fundamentally a data problem: no single party can see true demand, everyone reacts to the distorted echo of the party downstream, and the distortion amplifies as it travels up the chain.
The fix is not smarter forecasting at any one node — it is a shared view of the real signal, which no one in a fragmented, informal chain currently has. Name that for the founders in the room: the company that lets every link see the same demand truth is a real category, and it does not exist here yet.
Signal, amplified up the chain
Real consumer demand+5%
Order seen at the farm+40%
CauseEach link reacts to the echo, not the source