Helping students find the right courses

The product problem

Students browsing a large course catalog were not consistently finding relevant next steps. Low satisfaction and drop-offs affected enrollment, while universities needed more value from the platform and better visibility into recommendation performance.

The product goal was larger than showing a ranked list. Students needed a useful path through the catalog, advisors needed enough context to support those choices, and the platform needed measurable evidence that recommendations were improving the enrollment experience.

I treated drop-offs after recommendation-driven enrollment as a counter-signal. Higher clicks alone would not prove value if students abandoned the course later, so the product scorecard connected discovery, selection, enrollment, satisfaction, and retention.

The decisions

As Principal Product Manager, I shaped the goals, requirements, roadmap, and evaluation approach across academic product, data science, machine learning, architecture, UX, legal, security, and university stakeholders. I turned a broad AI opportunity into a staged system with product and program gates.

We started with trust and precision. The initial experience showed fewer, higher-confidence recommendations rather than maximizing the number of items. For new students and sparse histories, we used safe rules and richer signals rather than waiting for a perfect behavioral history. As the data foundation improved, the design could use richer model signals without losing the fallback path.

The delivery system

The work covered data readiness, model design, offline evaluation, human review, online evaluation, telemetry, and release gates. I partnered with data science and ML to improve offline precision by 50% and recall by 30%. We used A/B testing and interleaving to compare model behavior and improve relevance.

Human review and responsible AI checks were part of the lifecycle, with privacy and compliance treated as product requirements. The recommendation flow also enforced practical guardrails: eligibility prerequisites, no repeat courses, a rule-based fallback when model output was unavailable, and monitoring for enrollment behavior and model drift. Those controls made the user experience safer while giving the delivery team clear release criteria.

The outcome

The recommendation and personalization system increased engagement by 35%, reduced drop-offs by 50%, and increased conversion by 15%. Model relevance improved by 20% through online evaluation. The work also gave administrators a way to understand recommendation performance and gave the product a stronger value story for university customers.

The lesson is that AI product management is a system of decisions. A useful recommendation is only one part of the outcome; the experience must also be explainable enough to trust, safe enough to deploy, measurable enough to improve, and economical enough to support.