The track is [reference]: read, not yet practised. Operations research validates against a decision someone actually has to live with, not against a textbook answer. These projects are ordered from cheapest to most real. Each one says what “done” means.

Difficulty: 🟢 starter (hours–days) · 🟡 intermediate (a week or two) · 🔴 ambitious / portfolio-grade. ⭐ = highest signal.

Python + SciPy (HiGHS) is the whole toolchain. For anything beyond a toy, move to a modelling layer — PuLP or Pyomo — which wraps the same open-source solvers. Nothing here needs licensed software.

Projects

  • 🟢 Reproduce every worked example from memory — the workshop LP, the simplex trace, the branch-and-bound tree, the allocation DP, the M/M/2 comparison. Done when: you can rebuild each with the note closed and your numbers match. Exercises: notes 04, 07, 08, 12.

  • 🟢 ⭐ Model a decision you actually make — your weekly study timetable (subjects, hours, one subject per day, deadlines), or a week of meals on a budget meeting nutrition targets. Write it as an LP or MILP, solve it, and use the answer for a week. Done when: you’ve followed the plan, written down what the model got wrong, added the missing constraints, and re-solved. That loop — model, act, find the missing rule — is the discipline. Exercises: formulation, integer programming.

  • 🟢 A capacity calculator for a real service — take a service you’ve built (or a realistic invented one), measure or estimate arrival rate and service time, and use Little’s law and M/M/c to predict queueing at 1×, 1.5× and 2× load. Then load-test it and compare. Done when: you have a table of predicted vs measured latency and can explain the gaps (variability, warm-up, non-Poisson arrivals). Exercises: queueing. Ties to backend.

  • 🟡 ⭐ A rota generator — staff, shifts, availability, minimum cover per shift, maximum shifts per person, no back-to-back late-then-early, a fairness objective. Build it in PuLP or OR-Tools with a small web or CLI front end. Done when: it produces a valid rota for a realistic month in seconds, explains infeasibility in plain words (“Tuesday night can’t be covered because…”), and someone who does rotas by hand agrees the result is usable. Exercises: integer programming.

  • 🟡 A delivery route planner — a vehicle-routing problem with capacities on real addresses (geocoded, with road distances from an open routing engine). Compare nearest neighbour, 2-opt, and OR-Tools’ routing solver, against a lower bound. Done when: you have a table of method vs distance vs run time over several instances and seeds, with an honest gap estimate. Exercises: heuristics.

  • 🟡 Simulate a real queue — a clinic, a restaurant, a bus stop, a support inbox. Collect real arrival and service times (even 50 observations), fit distributions, simulate, and test one change (“add a second server at lunch”). Done when: your simulation reproduces the observed average wait within its confidence interval before you test the change. Exercises: simulation.

  • 🔴 Optimise, then simulate — a small supply chain: two warehouses, ten shops, random daily demand. Use a MILP to decide which warehouse serves which shop and the reorder points; then simulate a year of random demand against the plan and measure stockouts and cost. Iterate. Done when: you can show the plan’s simulated service level and cost, and one change the simulation caught that the optimiser couldn’t. Exercises: notes 07, 13, 14.

If you only do one

Model a decision you actually make, and live with the answer for a week. It’s the only project here where you’re also the client — and the client is where operations research is really tested.

↑ Back to Operations Research