MSc research · University of Sharjah

Proposal defended, November 2026. Analysis in progress.

Predicting late orders before they are late

Thesis title Using Machine Learning and Agentic AI to Measure, Explain and Predict Order-to-Delivery Lead Time in a Configure-to-Order Manufacturer

In plain words: find where the order time goes, then warn planners before an order runs late.

There are no thesis results yet. The demo runs the same method on synthetic data.

The problem

A customer who orders a built-to-spec product cares about two things. Will it arrive on the date they were given, and how long will they wait? That wait runs from the order date to delivery, through sales, engineering, planning, production, inspection and dispatch. The case is a configure-to-order manufacturer: it builds each product from options the customer picks. It quotes lead time from a standard table that leaves out part of the wait. Nobody measures the full wait, and nobody can tell at order entry if an order will be late. So the study puts reliability first and speed second. A shorter quote does not help if it is then missed.

A. The customer’s waitOrder date to delivery
B. What the quote countsClearance to ex-factory

Engineering and approval time the quote leaves out

Two clocks for one order. Clearance is the day the order is approved to build. Ex-factory is the day it leaves the factory. Not to scale.

Three questions

  1. RQ1

    Where does the time go?

    Split the wait into stages. Find the few stages that carry most of its length, spread and delay.

  2. RQ2

    Is the promise kept?

    Check how often the first promised date is met. Compare the promised interval with the wait the customer goes through.

  3. RQ3

    Can we see lateness coming?

    Predict the delivery date and the late risk on the day the order is entered. Compare with today’s promise. Use it to warn early and to target fixes.

How the study works

Four steps. Each one feeds the next. RQ1 to RQ3 are the three research questions above.

  1. MeasureRebuild every order line from the company’s order records and time each stage.RQ1
  2. ExplainFind the one or two stages behind most of the delay. Check the promise.RQ1 · RQ2
  3. PredictScore late risk and the delivery date at order entry, with data known that day.RQ3
  4. WarnAn agent watches open orders, flags likely delays and says why. People act.RQ3

Method in brief

A single case, studied with numbers. Everything rests on the company’s order records, at the level of the order line.

Data

Read-only queries on the company’s order records. The live system is never written to.

  • Several years of order lines
  • Made-to-order lines, linked to a production order
  • Cancelled, service and test lines left out

Measure and explain

Medians and percentiles carry the analysis, because a few orders wait far longer than the rest.

  • Median, IQR (the spread of the middle half), P90 (the slowest one in ten) and CV (spread compared with the average) by stage
  • Pareto of excess late days: the few stages behind most of the delay
  • First promise against actual delivery
  • Bootstrap confidence intervals: the data is resampled many times to show how sure each number is

Models

A decision tree first, because planners can read its rules and push back. A random forest shows what that readability costs.

  • Train on older orders, test on the newest ones
  • Only data known on the day of entry
  • Benchmark: today’s promise
  • Built in Python, repeated in RapidMiner as a cross-check

Early-warning agent

A small Python agent scores open orders each morning and notes any stage held longer than usual. A language model writes the alert in plain words. It has no write access. It flags; people act.

  • Tested by replaying the last year day by day
  • It sees only what was on file each morning
  • Counts late lines caught, how early, and false alerts

What-if scenarios

A simulation fed by the stage times tests how on-time delivery and lead time respond when a key stage is cut or steadied.

  • Example: one extra engineer in peak season
  • Used only once it reproduces the lead times on record

Validity and ethics

  • Read-only access to company records. No people are interviewed or surveyed.
  • The company stays unnamed and internal field names stay hidden.
  • Every duration is scaled by a factor that is not disclosed.
  • Older orders train the models and newer ones test them, so no future information leaks in.
  • Findings are reported as association, not cause.
  • The language model sees only anonymised fields, or runs on a set-up the company signs off.
  • Whatever the model says, the planner still owns the promise.

Stages of the wait

StageFromToYears
Pre-productionOrder dateProduction startAll years
ProductionProduction startLast unit completedAll years
Post-productionLast unit completedLine fully deliveredAll years
Order handling and engineeringOrder dateRelease to planningRecent years
PlanningRelease to planningProduction order createdRecent years
Production queueProduction order createdProduction startRecent years

The three finer pre-production stages can be traced only for recent years.

Research plan

From proposal to submission.

Proposal defended, November 2026. Analysis in progress.

  1. October to November 2026Done
    Proposal

    Literature review finished. Proposal submitted and defended.

    Proposal defended
  2. November 2026Now
    Data

    Company approval letter. Pull and clean the order data.

    Clean dataset
  3. December 2026 to January 2027Next
    RQ1 and RQ2

    Stages, spread and Pareto. The promise check.

    RQ1 and RQ2 results
  4. February 2027Next
    Conference paper

    A paper on RQ1 and RQ2.

    Paper submitted
  5. February 2027Next
    Models

    RQ3: the decision tree and the random forest.

    RQ3 models
  6. March 2027Next
    Agent and scenarios

    Agent build and replay test. What-if scenarios.

    Prototype and demo
  7. April to May 2027Next
    Write-up

    Results, discussion and conclusions. Submit the thesis. Start a journal paper.

    Thesis and journal draft

Dates may shift with the programme calendar.

Interactive demo

See the method run, on synthetic data

A fictional company with about 6,000 synthetic order lines. Time the stages, enter an order and watch the tree score it, replay a year of early warnings, and try two what-if levers.

Synthetic data for a fictional company. These are not thesis results.Open the demo