Pairwise v2
AI & Data
Semester programme:Complex Software Systems
Client company:Hedgecreek
Project group members:Niels van Lierop
Stijn van den Hurk
Keke Kusters
Yordan Markov
Project description
Pairwise v2 focused on improving an existing MVP into a more explainable and reusable matching engine. The main challenge was to make match results easier to understand, because the original version gave scores without enough insight into why a result scored high or low.
The project researched and developed a system that combines semantic matching, weighted criteria, feedback, monitoring, exports, and result explanations. The goal was to make the process more transparent, maintainable, and useful for future development.
Context
The project belongs to the domain of data-driven decision-support software and intelligent matching systems. Pairwise compares organizations or contacts against ideal customer profile criteria and ranks them based on structured data, semantic embeddings, and weighted scoring.
The original MVP already had working functionality, but it behaved too much like a black box. Therefore, the project focused on explainability, feedback learning, visualization, monitoring, and maintainable software architecture.
The technical context included a protected dashboard, backend API, database, vector search, embedding services, match history, exports, feedback events, and monitoring data. Because the project was developed under an NDA, confidential business details, internal data, and stakeholder-specific information are not included.
Results
The project delivered a working Pairwise v2 prototype with implemented core flows. The main outcomes were a protected dashboard, backend API, organization and contact matching, ICP quick matching, exports, match runs, feedback storage, and monitoring.
A key result was the improved pipeline. The system retrieves candidates using embeddings and vector similarity, then re-scores them using weighted criteria. This makes the results more flexible while still keeping them explainable. Match explanations help users understand why a result appears instead of only seeing a score.
The team also delivered a feedback foundation, monitoring overview, and handover documentation. These outcomes make it easier for future maintainers to understand, run, assess, and continue the project.
The project can be positioned around TRL 5 to TRL 6. It is more than a concept because the core flows were implemented and demonstrated in a relevant environment. However, it is not fully production-ready yet. Further work is still needed on access control, automated testing, CI/CD, secrets management, backups, and rollback procedures.
About the project group
Our project group consisted of four ICT students following the Complex Software Systems semester. We had different focuses, including backend development, frontend development, stakeholder communication, task management, data handling, and research.
The project lasted around 18 weeks and was developed through sprint-based work. We worked together on location during group days and used online communication during home-working days. Tasks were divided based on expertise and learning goals, while important decisions were discussed as a team and validated through feedback.