Smart Course Recommender - AI Product
Company: AECC Global | Role: PM - AI & Growth | Year: 2024
An AI-powered course recommendation engine built to personalise destination and course selection for international students at the point of counsellor engagement.
The Problem
AECC Global counsellors handle hundreds of student inquiries daily. Each student has unique goals, academic backgrounds, budgets, and destination preferences - but counsellors were manually triaging and recommending courses with no system support. Pain points discovered in discovery:
- Counsellors spent 20–30 minutes per student session on course shortlisting
- Recommendations were inconsistent across counsellors - heavily influenced by individual familiarity
- Students received a narrow set of options, often missing well-suited alternatives
- No feedback loop: poor recommendations weren’t surfaced or corrected The business impact of status quo:
- Lower conversion from meeting to application (counsellor fatigue + poor fit)
- High re-engagement rate as students came back with questions on alternatives
- Counsellor satisfaction declining - onboarding new counsellors took 4–6 weeks
Discovery Process
What I did
- Shadow sessions - sat with 8 counsellors across Bangalore, Chennai, and Hyderabad over 2 weeks to observe real consultations end-to-end
- Data audit - pulled 6 months of CRM data to understand patterns: which courses converted highest, where in the funnel drop-offs happened, top destination clusters
- Counsellor survey (n=45): ranked their biggest time sinks and confidence gaps
- Competitor benchmarking: IDP, AECC’s closest comp, had basic filtering - not personalisation
Key Insights
- Counsellors trusted “social proof” heavily - “other students from your background chose this” was a powerful framing
- There were also Entry Requirements specific to each source country and university to destination country
- Students wanted to see ranked options with reasons, not a flat list
- The CRM had rich historical data that wasn’t being used at all
Opportunity Sizing
| Metric | Current | Target |
|---|---|---|
| Time per course shortlist | 20–30 min | 5–10 min |
| Courses shown per student | 3–5 | 8–12 (ranked) |
| Counsellor confidence score | 6.2/10 | 8+/10 |
| Recommendation-to-application rate | ~34% | 45%+ |
Solution Design
Approach
A hybrid recommender combining:
- Collaborative filtering - “students like you chose these” (profile similarity on academic background, budget, destination preference)
- Content-based filtering - match course attributes to students requirements and eligibility criteria
- Rule-based boosting - counsellor-configured boosts for AECC's commercial partnerships and intake availability
What I wrote in the PRD
Core user story: As a counsellor, I want a ranked shortlist of courses generated from the student’s profile so that I can spend my session time on consultation, not manual searching. Non-goals I explicitly scoped out:
- We are NOT building a self-serve student recommendation tool (in v1)
- We are NOT replacing counsellor judgment - the tool augments, not automates
- We are NOT integrating live intake data in v1 (too complex, future release)
Key Design Decisions
Decision 1: Counsellor-first, not student-first (v1) Rationale: Counsellors are the primary conversion lever. Embedding recommendations into their workflow (Zoho CRM) drives adoption without changing student journey. Built for where users already live. Decision 2: Show the “why” behind each recommendation Rationale: Discovery showed counsellors won’t trust a black box. Each recommendation shows 2–3 reasons (“High acceptance rate for your profile”, “80% of similar students chose this”). This built trust and adoption. Decision 3: Feedback loop from day one Rationale: A thumbs-up/down on each recommendation creates the training signal for model improvement. Without this, the model stagnates.
Delivery
Timeline: 14 weeks (discovery to v1 launch)
| Phase | Duration | Key Output |
|---|---|---|
| Discovery & framing | 2 weeks | Problem statement, opportunity sizing |
| PRD + design | 3 weeks | Spec, wireframes, data model |
| Engineering (ML + BE + FE) | 7 weeks | Model, API, CRM integration |
| Pilot (2 offices) | 1 week | Feedback, bug fixes |
| Full rollout | 1 week | Training, comms, monitoring |
Team: 1 ML engineer, 2 backend engineers, 1 frontend engineer, 1 designer, 1 data analyst
Results
- Counsellor shortlisting time reduced from ~25 min to ~7 min per student
- Counsellor confidence score: 8.3/10 (vs 6.2 baseline)
- 1.4× increase in number of courses shown per student
- Model accuracy (thumbs-up rate): 74% at launch, 81% after 6 weeks of feedback loop
What I Learned
- Trust is the real adoption blocker for AI tools - explainability (showing why) was more impactful than accuracy improvements
- Launching in 2 pilot offices first gave us rich feedback before scale - we caught a critical bug in the profile-matching logic that would have tanked trust