A starting point the learner can question

Help a new learner choose a course that fits their goal, available time and experience. Explain why each course fits, and show when none qualifies. The suggestions use fixed rules.

The product choice: Explain the recommendation and give the learner control over the limits. Try the sample · Follow the walkthrough.

Experience and decision

My Oracle work included course recommendations, cold-start decisions and evaluation with data science and ML partners. This example isolates the product policy around relevance, eligibility and learner trust. This is a separate fictional portfolio exercise; its records, rules and results do not reproduce an employer system.

Decision to challenge: When should the product show no recommendation instead of a weak one? A strict match rule protects trust but can also hide a useful adjacent course; that tradeoff needs learner evidence.

User and decision

A new learner has no tracked behavioral history and sees no reason to trust an unexplained course list. Mo's product direction asks what can be useful and safe with little evidence. This original Northstar prototype answers with declared preferences, eligibility before ranking, inspectable reasons and reversible control. AI assisted implementation and testing; it is not an employer implementation, proprietary ranking system or customer study.

Prioritization and tradeoff

The main decision is to require a positive stated-goal match as well as duration, level and prerequisites. With a data goal and only 20 minutes, a 20-minute idea-sketch course has match 0/3. Filling that screen would imply relevance without evidence. The prototype instead explains no match and offers constraint editing and the whole catalog. Browsing remains a viable alternative for a learner who prefers to inspect everything.

Goal fit favors the strongest declared match; exploration favors authored breadth within the same boundaries. The explanation drawer shows the exact calculation and reasons for qualification or exclusion. Immediate not-for-me feedback excludes only the named course, can be restored and does not imply model retraining. Preference and reset changes are reviewed before saving. A changed local scenario requires a fresh review, protecting the preferences actually inspected.

Demonstrated outcome and limits

The independently calculated fictional fixture has four eligible courses, three judged useful. Goal fit's top three covers 3/3 useful eligible courses and includes 3/3 useful shown; exploration covers 2/3 and includes 2/3. Both contain 0/3 rejected shown. Exploration introduces the neutral ethics course. The rejected outline is excluded by zero goal match. These figures describe one authored persona, eight authored judgments and a chosen top-three cutoff; they cannot establish real relevance, engagement or commercial performance.

Product tradeoffs remain provisional: the positive-match threshold may overexclude useful adjacent learning; breadth scores require learner evaluation; declared completions can be inaccurate; one-step Undo is simpler but offers limited recovery. Decisions and risks records alternatives. Validation separates software evidence, publication gates and proposed human research. No research, personal review or commercial outcome is claimed.

Next investment decision

The investment question is whether declared preferences and inspectable reasons help a new learner choose a useful next course compared with browsing the catalog. Use the proposed 5–8-person study to challenge the positive-match threshold and authored breadth judgments: a useful adjacent course may be excluded even when every rule works. If explanations add effort without improving choice, simplify them; if no-match recovery fails, revise the constraint workflow before adding ranking complexity. A future learned policy would need independently collected relevance judgments and evidence of useful learning progression relative to its operating cost. The authored fixture cannot justify that investment.