01 / PROFILE
Collect useful signals
The questionnaire turns student context into structured recommendation inputs.
AI Project Recommender
Recommendations students can inspect, not a chatbot answer they are expected to trust blindly.
- ROLE
- Product strategy · Interface design · Full-stack engineering
- STATUS
- Thesis product
- STACK
- React · Node.js · Express · MongoDB · Gemini
PRODUCT JOURNEY / 03THE PRODUCT,
THE PRODUCT,
IN USE.
One complete interface state at a time. Each screen documents the task, the system response, and what the user can do next.
02 / SCORE
Build an auditable ranking
Deterministic criteria rank candidate projects and expose fit signals.
04 / MEASURE
Inspect recommendation quality
Feedback and run analytics make recommendation behavior reviewable over time.
PROFILE → SCORE → PERSONALIZE → MEASURE
WHY THIS PRODUCT
NEEDED TO EXIST.
Generic project lists ignore skill fit and difficulty. Pure LLM recommendations can sound convincing while remaining unstable, difficult to audit, and useless during provider failure.
I built the recommendation experience around structured profile signals and deterministic scoring first. Gemini adds personalized reasoning and reranking, while a fallback set keeps the product functional when the model is unavailable.
THE OPERATING
SEQUENCE.
- 01PROFILE
Collect useful signals
The questionnaire turns student context into structured recommendation inputs.
- 02SCORE
Build an auditable ranking
Deterministic criteria rank candidate projects and expose fit signals.
- 03PERSONALIZE
Add contextual reasoning
Gemini reranks or explains within a controlled boundary rather than inventing the whole result.
- 04MEASURE
Inspect recommendation quality
Feedback and run analytics make recommendation behavior reviewable over time.
DECISIONS THAT
SHAPED THE BUILD.
Structure before prompting
Skills, interests, difficulty, goals, and constraints become comparable signals rather than one vague free-text request.
Deterministic first
A visible scoring layer produces the base ranking; AI can personalize the reasoning without erasing why a project matched.
Failure still returns value
If Gemini is overloaded or unavailable, the scored recommendation set and explanations remain usable.
SUPPORTING
PRODUCT SURFACES.
Secondary screens that extend the core journey without diluting it.
05 / 06Project brief
06 / 06Positioning
05 / MY RESPONSIBILITYSCOPE OF
SCOPE OF
RESPONSIBILITY.
I led the product concept, scoring architecture, recommendation flow, result explanations, analytics interface, backend integration, and the thesis-level product presentation.
The system produces ranked projects, fit summaries, score breakdowns, recommendation explanations, feedback signals, run analytics, and exportable results with a non-LLM fallback path.
- Hybrid recommendation
- Explainability
- Fallback design
- Analytics