Every Monday, the Prediction Engine gives each of Upokul’s eleven hundred field officers up to fifteen borrowers who are likely to fall behind within six weeks, with the reasons in Bangla. That list is useful only if the visits happen, and happen in the right order. An officer’s week is already mostly full: group meetings take four mornings; visits fit into afternoons and one flexible day; travel is by motorcycle, rickshaw van and, in the monsoon, by boat; and roads that were open last week may be under water this one.

Left to themselves, officers do what any sensible person does. They visit borrowers near the morning’s meeting point, or borrowers they know well, or whoever is easiest to reach. Branch managers help by dividing the list by village. None of that is wrong. It simply ignores the thing that matters most: which visit, this week, protects the most at-risk balance for the time and distance it costs.

The scale is typical of the sector. Bangladesh’s licensed microfinance institutions work through more than 26,000 branches and over 220,000 employees, and much of that workforce spends its days travelling between borrowers. Routing research has long shown that computerised planning typically saves 5 to 20 percent of transport costs. For an MFI, the larger prize is not fuel. It is officer hours arriving at the right door before a missed installment becomes two.

Formally, each officer’s week is a routing problem in which visits carry different values and not every flagged borrower can be reached. Across ninety branches and eleven hundred officers, that is far too large to solve exactly on a Sunday night. It is precisely the case for heuristics - provided the plan says how good it is.

Sunday night, a plan per officer

The head of operations asked:

> Plan next week’s afternoon visits for every field officer. Maximise the at-risk balance we reach before it falls into arrears, within each officer’s free hours and current travel times, visiting each flagged borrower at most once. Keep officers inside their own branch area.

The Optimization Engine used that last sentence to decompose the problem into ninety branch-level problems that can be solved side by side. Decide: which flagged borrowers each officer visits, on which afternoon, in what order. Value of a visit: outstanding balance, multiplied by the Prediction Engine’s probability of two missed installments, multiplied by the effect of an early visit. The engine flagged that last factor as an assumption: Upokul has no controlled evidence yet on how much a visit helps, so the plan uses a cautious uniform effect and states that in the specification. Must hold: officer hours after group meetings, travel times on the current road network with flood closures and boat legs reported by branches, and one visit per borrower.

Each branch problem ran a heuristic search for about two minutes, in parallel, while a relaxation calculated a bound on the best possible plan. Every branch plan therefore arrived with a gap: the most value it could possibly be leaving on the table. Most branches came in under two percent. Four were above four percent, all dependent on ferry timetables that weaken the bound. The engine flagged those branches instead of averaging them into a reassuring headline.

At-risk balance reached per officer, per week

1.65 lakh taka1.52 lakh taka2.17 lakh taka

Kilometres travelled per officer, per week

62 km51 km53 km

Officer's own choiceNearest firstOptimization Engine plan

+31% of at-risk balance reached, 14% less travel. Officers can swap any visit; the rest of the week re-plans around it.

Composite scenario; figures illustrative. Travel times include flood closures and boat legs reported by branches.

Compared with officers’ own choices, the plans reached about thirty-one percent more at-risk balance per week with fourteen percent less travel. The officers’ knowledge still matters. Any visit can be swapped in the tablet app - a borrower who is away at her mother’s village, a bridge the flood report missed - and the rest of that officer’s week re-plans around the change. Local knowledge enters as a correction to the plan rather than a reason to ignore it.

Branches

010203040

3431214

under 1%1–2%2–4%4–7%

Gap between the plan found and the proven bound

• Every branch plan ships with its gap
• The four above 4%depend on ferry timetables, which weaken the bound
• Heuristic searchabout 2 minutes per branch, in parallel

Composite scenario; figures illustrative. Bounds from a relaxation of each branch's visit-planning model.

Where the loan fund goes before planting

The second question is seasonal. Before the Boro planting season and again before Eid, disbursement demand rises sharply across the coastal branches. The loan fund - from apex lenders, debt funds and members’ savings - is finite. Funder covenants limit how concentrated the portfolio can become in agricultural lending. Branches above an internal portfolio-at-risk threshold receive capped new disbursement. Cash transfers to each branch are limited by security rules.

The managing director asked for the allocation that meets the most eligible demand over the next four weeks within all of those rules. This model is small enough to solve exactly, and it was proven optimal. More useful than the allocation was the engine’s report on binding constraints. The agricultural concentration covenant bound first, and the engine calculated what it cost: each additional percentage point of allowance would have funded input loans for roughly nineteen hundred more eligible borrowers.

The managing director took that figure to the funder. The funder agreed to a temporary seasonal allowance with a monitoring condition, reported monthly through the same Analytics Engine layer that had supported the last facility renewal. It was a negotiation grounded in a computed number rather than a request for flexibility.

What the three engines add up to

Upokul now works through the full chain. The Analytics Engine produced exposure to the taka thirty-five minutes after signal seven. The Prediction Engine finds borrowers weeks before the second missed installment. The Optimization Engine decides which doors each officer reaches and where the loan fund goes when demand exceeds it.

The rules remain human. A score or a plan can trigger a visit and a conversation; it cannot trigger a penalty, refuse a loan or change a borrower’s terms. Officers and branch managers can override every plan, and every override is recorded. What the field officer does at the door - listening, adjusting an installment day, helping a household through the lean weeks - is still the work that matters. The plan simply makes it more likely to be the right door, in the right week.