TeamUp: Semantic Project Matching and Team Formation for Learning at Scale

Technical Summary: TeamUp – Semantic Project Matching and Team Formation for Learning at Scale

Problem Statement
Project-based learning (PBL) is a cornerstone of contemporary engineering education, yet scaling it to cohorts of 200–300+ students presents significant pedagogical challenges. Current allocation methods, relying on manual spreadsheets, ad hoc preference surveys, and instructor intuition, fail to address three persistent issues:

1. Poor Skill-Project Fit: Subtle competency differences are overlooked, leading to projects that are either too easy (causing disengagement) or too difficult (causing overwhelm), misaligning with the student's Zone of Proximal Development (ZPD).
2. Homogeneous Team Clustering: Manual or self-selected teams often group students with identical strengths, missing the cognitive diversity required for enhanced problem-solving and learning outcomes.
3. Inequitable Opportunity Distribution: Without principled mechanisms, high-performing students self-select visible projects, while underrepresented groups face reduced access to opportunities.

Traditional approaches treat allocation as a logistical optimization problem rather than a learning design problem, failing to capture the multidimensional nature of student competencies and project requirements.

Methodology
TeamUp is a lightweight, embedding-based system designed to automate project allocation and team formation while adhering to learning science principles. The system architecture integrates four core components:

1. Data Preparation & Embedding Generation:

   • Input: Students provide profiles including self-assessed skill proficiency, coursework, project experience, and domain preferences. Resumes can be uploaded for NLP-based extraction. Supervisors define projects with required/optional skills, difficulty ratings, and domain tags.
   • Vectorization: All textual and categorical data are converted into dense vector representations using pretrained language models (e.g., text-embedding-ada-002, Google Gemini).
   • Aggregation: Student embeddings ($s$) are computed via weighted averaging of skill and experience embeddings, where weights correspond to self-reported proficiency levels. Project embeddings combine required skills (1.5× weight), optional skills (0.75× weight), and descriptions.
   • Storage: Vectors are stored in PostgreSQL using the pgvector extension, enabling efficient similarity search via HNSW indexing without separate vector database infrastructure.

2. Similarity-Based Matching (Hybrid Ranking):

   • Base Score: Initial matching uses cosine similarity between student and project embeddings.
   • Pedagogical Adjustments: Raw similarity scores are modified by three constraints:
     • Difficulty Alignment (ZPD): A quadratic penalty is applied if the project difficulty significantly mismatches the student's experience level, preventing overwhelming or boring assignments.
     • Domain Preference: Scores are boosted (15–20%) if the project aligns with the student's stated domain interests, supporting self-determination theory.
     • Demand Balancing: Scores decay exponentially as a project's subscription ratio approaches capacity, ensuring equitable distribution of high-demand projects.

3. Team Complementarity Modeling:

   • To prevent homogeneous clustering, TeamUp employs a greedy algorithm that maximizes skill complementarity.
   • Starting with the highest-matching student, additional members are selected based on a complementarity score: $comp(t, c) = \alpha \cdot sim(c, p) - \beta \cdot sim(t, c)$, where $t$ is the current team's average embedding, $c$ is a candidate, and $p$ is the project.
   • This formulation balances project fit against penalizing candidates who are too similar to existing team members. Team diversity is quantified by the variance of the team-averaged embedding vector.

4. System Architecture:

   • Built with Next.js/React (frontend) and Supabase/PostgreSQL (backend).
   • Supports real-time recommendations with sub-second latency for cohorts of 500+ students.
   • Operational costs are maintained under $0.10 per student.

Key Contributions

1. Principled Allocation System: A semantic embedding-based system grounded in learning science (ZPD, self-determination theory, collaborative learning) that moves beyond simple keyword matching or preference ranking.
2. Cognitive Diversity Modeling: A novel approach to team formation that explicitly models skill complementarity through embedding variance, operationalizing the principle that diverse teams outperform homogeneous ones.
3. Responsible AI Design: The system is designed as a "human-in-the-loop" tool that provides transparent, explainable recommendations (showing skill alignments and rationale) while preserving instructor authority for final decisions. It operates within realistic educational budgets without requiring custom training data or specialized ML infrastructure.
4. Virtual Validation: Evidence from a virtual experiment demonstrating the system's ability to produce superior allocations compared to random assignment.

Results (Virtual Experiment)
The system was evaluated using 250 dummy student profiles and 60 project descriptions, comparing TeamUp against random allocation:

• Match Quality: TeamUp achieved a mean cosine similarity of 0.74 versus 0.43 for random allocation.
• Difficulty Alignment: 83% of TeamUp students were placed within one difficulty level of their experience, compared to 34% under random assignment.
• Team Diversity: 82% of TeamUp teams covered three or more technical areas, compared to 41% of random teams.
• Performance: The system generated all recommendations in 4.2 seconds (sub-second per query) with an API cost of $0.024 for the full cohort.

Significance and Claims
The paper claims that TeamUp offers a scalable, cost-effective solution to the pedagogical challenges of large-scale PBL. By leveraging semantic embeddings, the system addresses the limitations of manual allocation in capturing nuanced skill-project fits and fostering cognitive diversity.

The authors explicitly state that the current results are preliminary, derived from a virtual experiment with synthetic data. They acknowledge that dummy data cannot capture the "messiness" of real student profiles, self-assessment inconsistencies, or interpersonal team dynamics. Consequently, the paper does not claim that TeamUp guarantees improved learning outcomes or equity in practice yet. Instead, it positions the system as a promising design pattern for responsible AI in education, with the critical next step being a controlled deployment with real users to validate fairness, user satisfaction, and actual learning gains. The system is framed not as a replacement for human judgment, but as a tool to handle routine matching, freeing instructors to focus on edge cases and contextual decisions.