Skip to content

Team Formation at Scale

Forming 20–30 balanced teams from a pool of 114+ students is surprisingly difficult. As class size doubles, the number of possible team combinations grows exponentially — it's an NP-Hard partitioning problem. Random assignment ignores skills and preferences. Student self-selection produces cliques. Manual sorting by instructors doesn't scale.

Our solution: a Neuro-Symbolic pipeline that combines machine learning with formal optimization.

The Pipeline

Step 1: Student Survey (Psephos)

At the start of each semester, students complete a comprehensive intake survey via Psephos, our extensible survey platform. The survey captures: - Self-reported technical skills and experience levels - Preferred programming languages and tools - Work style preferences (pair programming vs. solo work) - Availability and scheduling constraints - Learning goals and project preferences

Survey data is managed through UMACS, our RBAC-based access control system, ensuring that sensitive student information is properly secured.

Step 2: Skill Extraction (SkillNet)

Raw survey data is noisy. A student who lists "Python" as a skill might mean they took one introductory course — or they might have built production systems.

SkillNet uses Graph Neural Networks (GNNs) to infer actual competency levels by: 1. Mapping self-reported skills onto the Computer Science Ontology (CSO) — a comprehensive knowledge graph of CS concepts and their relationships 2. Analyzing the relational hierarchies between skills (if a student claims "machine learning," do they also list prerequisites like linear algebra and statistics?) 3. Cross-referencing survey data with GitHub portfolio analysis — actual commit history, branching patterns, and technical contributions 4. Generating skill embeddings that capture both stated and inferred competencies

SkillNet functions as an MCP-enabled tool — it interfaces with the ontology and survey data through the Model Context Protocol, making it accessible to downstream systems.

Step 3: Constraint Optimization (OptiTeam)

With skill embeddings in hand, OptiTeam translates instructor goals into formal Integer Linear Programming (ILP) constraints:

  • Balance: Every team should have a mix of frontend, backend, and DevOps skills
  • Diversity: Teams should be diverse along relevant dimensions (experience, background, perspective)
  • Preferences: Where possible, respect student preferences for project topics and teammates
  • Constraints: No team of all experts, no team of all novices, minimum skill coverage for each project

OptiTeam uses a multi-agent LLM architecture: - A Manager agent interprets the instructor's natural-language requirements - A Formulator agent translates those requirements into ILP equations - An Evaluator agent checks the generated constraints for correctness and completeness

The system then solves for the global optimum — the best possible team configuration given all constraints simultaneously.

Step 4: The "Transfer-then-Adapt" Privacy Strategy

A key design challenge: how do you train a system to understand skill relationships without exposing student data to training pipelines?

The Transfer-then-Adapt methodology solves this: 1. SkillNet is initially trained on large-scale industrial datasets (GitHub open-source contributions, professional skill surveys) 2. The model is then fine-tuned on anonymized, de-identified educational data 3. Student data is never exposed to the foundational training process

Deployment Timeline

The pipeline transforms the first three weeks of the semester:

Week Activity
Week 1 HR Intake — Students complete the Psephos survey
Week 2 Deployment Memo — SkillNet processes survey data, OptiTeam generates team assignments
Week 3 IT Provisioning — Teams are announced, repositories are provisioned, projects begin

Empirical Validation

We evaluate team formation quality through a mixed-methods triangulation framework: - Subjective measures: Surveys (TAM, Self-Efficacy, NASA-TLX) capture student perceptions - Objective measures: MongoDB repository metadata (PRs, commits, CI logs) captures actual behavior - Correlation analysis: Pearson correlations detect Self-Report Bias — comparing reported peer review frequency against actual PR comment volume

Early results show that algorithmically-formed teams exhibit more balanced contribution distributions and higher satisfaction scores compared to self-selected or randomly-assigned teams.

This work has been presented at CSCI 2025. See the OptiTeam research page for the full technical description and the Empirical Triangulation Framework for the evaluation methodology.