SERVICES
│ Standalone course
Choosing algorithms for AI-enabling operations
Organize LLMs, supervised/unsupervised machine learning, Bayesian optimization, linear/integer programming, network flow, and rule-based approaches by INPUT/OUTPUT, runtime, and concrete examples. Build the criteria for "which to use for what."
─ ARTICLE COVER
System Design Skills ／ Standalone course
§I ─ POSITION
A course so you don't get the "technology choice" wrong when AI-enabling operations. Includes an AS-IS → TO-BE adaptation example for Azika Foods' ordering work and a phased-rollout design.
§II ─ TAKEAWAYS
Hold the criterion: "if the objective is to minimize/maximize, use optimization; otherwise, AI + human confirmation"
Speak, in one table, to each algorithm's INPUT/OUTPUT, runtime, and scaling characteristics
Design phased rollout (rule-based → ML → optimization) to lower risk
§III ─ EXCERPT
¶ 01
When AI-enabling operations, "an LLM can do anything" or "let's just use machine learning" is the most dangerous way to choose. The judgment axis is simple: "can the objective be quantitatively minimized/maximized?" If yes, an optimization algorithm; if no, ambiguity remains, so AI + human confirmation is needed.
¶ 02
Organize the technologies commonly used in operations into the following six cards.
| Card | Strength | Suited operations |
|---|---|---|
| Rule-based | Fully deterministic / explainable | Processing where the rules are clear and rarely change |
| Supervised ML | Needs labels / strong at prediction tasks | Demand forecasting / defect detection |
| Unsupervised ML | No labels needed | Clustering / anomaly detection |
| LLM | Free text / strong with ambiguity | Inquiry summarization / document classification |
| Bayesian optimization | Experiment design where trials are costly | Recipe optimization / hyperparameter search |
| Linear/integer programming | Exact minimum / maximum | Vehicle dispatch / delivery / scheduling |
¶ 03
Putting AI features in from the start is high-risk. Stepping through "rule-based → machine learning → optimization" lets you match the speed at which the field learns. At each step, make explicit as a requirement the conditions for advancing to the next (accuracy / utilization / field satisfaction).
¶ 04
Real example: the work of deciding order quantities. AS-IS is the chief's rules of thumb. Phase 1 makes it rule-based ("order quantity = recent 4-week average × safety factor"), Phase 2 introduces demand-forecasting ML, and Phase 3 develops into inventory-constrained optimization (integer programming). Line up each step's INPUT/OUTPUT, operating load, and expected effect to debate them.
¶ 05
In workflow design, the core of judging "where can be automated and where needs human confirmation" is just one thing: "can the technology used uniquely determine the solution?" This decides whether to put a human-check step into the business workflow.
¶ 06
Classify the technologies used in operations into four by the nature of the solution. You can decide which drawer to pull from before implementing.
| Type | Nature of the solution | Representative examples |
|---|---|---|
| Rule-based | Fully fixed | Inventory-threshold alerts, tax calculation |
| Optimization (exact solution) | Uniquely fixed | LP / MIP / network flow |
| Optimization (approximate solution) | Probabilistically best | Bayesian optimization, genetic algorithms |
| AI / ML / LLM | Probabilistic / ambiguous | Demand forecasting, text generation, image recognition |
¶ 07
When you use machine learning, accuracy won't be 100% and misjudgments always creep in. Rather than treating this as "the accuracy is bad," design the business workflow so it works even with a "human-check step" built in. Conversely, build a process whose solution is uniquely determined with machine learning, and after go-live you can't explain "why a different answer came out" and lose trust. Technology choice is one with workflow design.
§∞ ─ DOWNLOAD
─ EXERCISE
5 exercises: technology choice / AS-IS analysis / TO-BE workflow design / finding errors / adapting to your own work. Includes an exercise where you classify operational steps on the "is the solution unique?" axis and assign the appropriate technology.
Lecture slides
Every slide of the course
Cheat sheet
Quick reference for the field
Exercises
With answers and grading
Advanced / test
Hands-on plus rubric
INCLUDED PDFS ─ 4files
§IV ─ RELATED