BY DOMAIN, with what the AI actually does in each: HEALTHCARE medical IMAGE INTERPRETATION — tumours in radiology, diabetic retinopathy in retinal photographs; matches or exceeds specialist accuracy on narrow, well-defined tasks DIAGNOSIS SUPPORT from symptoms and history DRUG DISCOVERY — predicting protein structure and candidate molecules; AlphaFold's protein-structure predictions are the standout scientific result patient TRIAGE and readmission-risk prediction NOTE THE PATTERN: AI does the perceptual and statistical part, and a clinician retains the decision — because the cost of an error is high and accountability must rest with a person. TRANSPORT autonomous and driver-assist systems — perception, localisation, path planning, control ROUTE OPTIMISATION and traffic-signal control (the real-time example from the multiprocessor topic) predictive maintenance from sensor data FINANCE FRAUD DETECTION — the archetypal successful application, because there is abundant labelled data, a clear objective, and errors are recoverable credit scoring · algorithmic trading · anti-money-laundering REGULATORY CONSTRAINT: in many jurisdictions a declined credit application must be explainable, which rules out opaque models regardless of their accuracy NATURAL LANGUAGE machine translation · speech recognition and synthesis · summarisation · question answering · sentiment analysis · conversational assistants VISION face recognition · object detection · optical character recognition · industrial quality inspection · satellite and agricultural imagery INDUSTRY AND AGRICULTURE visual defect inspection on production lines · robotic assembly · yield prediction · pest and disease detection from leaf images · irrigation scheduling → the agricultural applications matter particularly in Nepal's context, where a phone camera plus a trained model can substitute for scarce extension officers GAMES — historically the field's proving ground chess (Deep Blue, 1997) · Go (AlphaGo, 2016) · poker (imperfect information) · real-time strategy WHY GAMES: the rules are exact, the objective is unambiguous, and progress is measurable — the properties real problems usually lack. EDUCATION adaptive tutoring and spaced repetition · automated marking · content recommendation A STUDY TRACKER IS AN INSTANCE: scheduling revision by how long ago a topic was ticked is the elementary form of the spacing effect, and predicting which topics a student will forget first is a genuine machine-learning problem. EXPERT SYSTEMS AND OPERATIONS diagnosis and configuration (ACtE0904) · scheduling · resource allocation · anomaly detection in infrastructure

THE CHECKLIST. A problem is a good candidate when it satisfies most of these, and the failures in practice are almost always a missing item rather than a weak algorithm. 1. IS THERE A CLEAR, MEASURABLE OBJECTIVE? "Reduce fraudulent claims" is measurable. "Improve customer happiness" is not, until someone defines a proxy — and the choice of proxy then determines what the system optimises, including its unintended behaviour. 2. IS THERE ENOUGH RELEVANT DATA, AND IS IT LABELLED? Supervised learning needs examples WITH ANSWERS. The usual blocker is not volume but labels: a hospital may have a million scans and no reliable record of which showed disease. 3. IS THE FUTURE LIKE THE PAST? A learned model assumes the distribution it was trained on continues to hold. Where behaviour shifts — a new fraud technique, a pandemic, a changed regulation — the model degrades. This is DISTRIBUTION SHIFT, and it is why deployed models need monitoring and retraining rather than one-time delivery. 4. IS THE COST OF AN ERROR TOLERABLE, AND IS THE ERROR RECOVERABLE? A wrong film recommendation costs nothing. A wrong braking decision is unrecoverable. THIS ONE DECIDES THE ARCHITECTURE: high-cost irreversible decisions need a human in the loop, a conservative default, or both. 5. CAN THE SYSTEM'S DECISIONS BE EXPLAINED, IF THEY MUST BE? Where law or professional duty requires a reason, an unexplainable model is unusable however accurate. 6. IS A SIMPLE RULE GOOD ENOUGH? THE MOST NEGLECTED QUESTION. If a hand-written rule achieves 90% of the benefit, it is usually the right answer: no training data, no drift, no infrastructure, and it can be read and audited. A WORKED COST CALCULATION — why accuracy alone never settles it. A fraud detector on 200,000 claims a year, 1% genuinely fraudulent (2,000 cases), average fraud loss NPR 40,000. The model has 95% recall and 90% precision. true positives = 2,000 × 0.95 = 1,900 caught false negatives = 2,000 − 1,900 = 100 missed false positives: at 90% precision, of every 100 flagged, 10 are innocent. Flagged total = 1,900 / 0.90 = 2,111, so false positives = 2,111 − 1,900 = 211 VALUE OF FRAUD CAUGHT 1,900 × 40,000 = NPR 76,000,000 LOSS STILL INCURRED 100 × 40,000 = NPR 4,000,000 COST OF INVESTIGATING FALSE ALARMS 211 × 2 hours × NPR 500/hour = NPR 211,000 NET BENEFIT ≈ 76,000,000 − 211,000 = NPR 75.8 million THE MODEL IS CLEARLY WORTH DEPLOYING. NOW CHANGE ONE NUMBER: suppose a false positive means a legitimate applicant is refused and complains publicly, at a reputational and remediation cost of NPR 100,000 each: 211 × 100,000 = NPR 21,100,000 net benefit falls to ≈ NPR 54.9 million — still positive. AND AT 60% PRECISION instead of 90%: flagged = 1,900/0.60 = 3,167, false positives = 1,267 1,267 × 100,000 = NPR 126,700,000 NET BENEFIT IS NOW NEGATIVE (−NPR 50.7 million). THE LESSON: THE SAME RECALL WITH DIFFERENT PRECISION FLIPS THE DECISION, and which metric matters depends entirely on the relative cost of the two error types. A model reported as "95% accurate" tells you almost nothing without the confusion matrix and the costs. THE BASE RATE TRAP, which the above quietly contains: with only 1% fraud, a model that flags NOTHING is 99% accurate. ACCURACY IS A USELESS METRIC ON IMBALANCED DATA — a point that returns in the machine learning section.

WHAT CURRENT AI DOES BADLY — stated honestly, because a graduate who can name these is more useful than one who cannot: COMMON SENSE AND CAUSATION Systems learn CORRELATION. They do not, by default, know that pushing a glass causes it to fall, and they cannot reason about what would have happened otherwise. BRITTLENESS OUTSIDE THE TRAINING DISTRIBUTION Performance can collapse on inputs that look trivially different to a human. ADVERSARIAL EXAMPLES — imperceptible pixel changes that flip a classification — demonstrate that the learned function is not the one we assumed. EXPLAINABILITY Large models cannot say why. Post-hoc explanation techniques give plausible accounts that are not guaranteed faithful to the actual computation. DATA HUNGER AND COST Training large models requires data and computation on a scale that concentrates capability in a few organisations. CONFIDENT ERROR A model's confidence is not calibrated to its correctness. Language models produce fluent false statements ("hallucination"), and fluency is exactly what makes the errors hard to notice. THE RISKS THAT ARE PROFESSIONAL RESPONSIBILITIES, not philosophical speculation: BIAS AND FAIRNESS A model trained on historical decisions reproduces the historical pattern, including its discrimination. If past lending favoured one group, a model trained on it will too — and will do so consistently, at scale, and with the appearance of objectivity. THE APPEARANCE OF OBJECTIVITY IS THE HARM: a biased human can be argued with, while "the system decided" ends the conversation. MITIGATION: audit outcomes by group, not just overall accuracy; examine what the training data represents; document the intended use. PRIVACY Models can memorise and reveal training data; aggregated data can re-identify individuals. AUTOMATION AND WORK Task automation displaces specific tasks rather than whole jobs, but the distribution of who gains and loses is a real social question rather than a technical one. ACCOUNTABILITY When an automated decision harms someone, who is responsible? "The model" is not an answer a legal system accepts, and designing for accountability means keeping a human decision point where consequences are serious. SECURITY models can be attacked — poisoned training data, adversarial inputs, prompt injection, model extraction THE PROFESSIONAL POSITION WORTH ADOPTING: an engineer building these systems is responsible for asking whether the problem should be solved this way, not only whether it can be. That is the same obligation the software engineering unit's requirements-validation activity describes — checking you are building the right thing — and it applies with more force when the system's decisions affect people who never agreed to be measured by it.