DuoAI
Many Azerbaijani high school students struggle to choose the right university major because they lack access to personalized career guidance. Important decisions are often influenced by family expectations, entrance exam scores, and limited knowledge of different majors and career opportunities. CareerPath AI addresses this problem by providing an accessible, AI-powered career guidance platform in Azerbaijani. Through a 12-question interest questionnaire based on the Holland RIASEC model, students receive a personalized interest profile and their top three matching university majors, along with explanations and potential career paths. An AI chat advisor also allows students to ask follow-up questions in their own language. Our goal is to help students better understand their interests, explore suitable educational opportunities, and make more informed decisions about their future.
Preliminary score
- Value for the user15.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI15.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing10.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility7.5 / 15
Data requirements, running costs and a clear next step.
- Originality4.5 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary49/100
- Strengths
- Your split between rule-based RIASEC cosine-similarity matching and an LLM used only for follow-up conversation is a sensible design that keeps the rankings predictable and explainable. Your testing write-up is honest and specific. You diagnosed why all-1 and all-5 profiles produce identical rankings (cosine similarity ignores magnitude), and you showed that match scores cluster between 94% and 100%. Your screenshots show a working end-to-end flow in Azerbaijani, and you are open about the 23-major dataset being your own estimate rather than official psychometric data.
- To improve
- Your repository README is still the original Lovable prompt, which says to 'use mock data for the results'. The Java Spring Boot backend appears in your deck but not in your components list or code setup, so please document what actually runs and where. Your 7-profile test has no stated expected outputs, so turn it into a table of input profile, expected majors and actual Top 3, and add even a small comparison with a counselor or students' current choices. Measure the cost per chat session and name the OpenRouter model you use. Also tie the pitch to your track, for example a school or university admissions office adopting it with a named metric.
- GPT-6 AstraPreliminary57/100
- Strengths
- Your project addresses a clear local need with a concrete outcome: three majors students can explore through Azerbaijani-language follow-up questions. Your questionnaire and results screenshots support the proposed workflow, and you clearly distinguish rule-based matching from the LLM’s conversational role. Your testing identifies specific cosine-similarity limitations and appropriately avoids treating seven successful profiles as proof of broad accuracy.
- To improve
- Your prototype evidence would be stronger with a recorded questionnaire-to-results-to-AI-answer run, reconciliation of the README’s mock-data instruction, and exact model and backend configuration. Document test inputs, expected rankings and actual outputs, then compare shortlist usefulness against students’ current advice or search process. Validate your estimated major vectors with reliable sources or counselors, explicitly handle uniform profiles, and label similarity percentages so students do not mistake them for validated suitability probabilities. Before your pilot, measure Azerbaijani answer accuracy, response time and per-session operating cost.
The team
Gülnar İsmayılovaLead
data
Xədicə Zeynalova
developer