Proprietary Technology and Governance Framework
TL;DR
ProFinda’s platform uses proprietary AI and a sophisticated ontology (90,000 skills, 900,000 terms, 15 million interconnects) to efficiently match workforce supply and demand. In line with the EU AI Act, they’ve identified all high-risk AI applications (Profile Match, CV Parser, Generative Role Descriptions, and Skills Suggestions) and built in specific, layered governance to ensure a low total risk.
ProFinda continuously leverages the latest AI and machine learning models across its platform to match workforce supply and demand efficiently. Recognizing the technological risks inherent in new AI models, ProFinda is fully committed to compliance with regulations like the EU AI Act and its conformity assessment requirements.
This paper details ProFinda’s proprietary technology and governance framework, demonstrating how high-risk AI applications – including the CV parser, Generative AI Role Descriptions, Profile Matching, Skills Framework and Skills Suggestions are engineered from the ground up to minimize bias, ensure traceability, and maintain human oversight. ProFinda’s proprietary global and local ontology and specialized AI architecture ensure a low total risk assessment across all high-value applications.
1. The ProFinda AI Foundation
ProFinda’s primary goal is to match resource supply and demand in the most efficient and effective way possible. This is achieved through proprietary matching algorithms that identify people with the optimal combinations of skills and experience for a project demand. The process relies on a multi-stage Natural Language Processing (NLP) pipeline that reads and understands human capital data to build the necessary profiles for the matching algorithms.
Built for Auditing, Traceability, and Bias Mitigation
Years of research and iterative refinement have led to the development of services rooted in data governance, security, and ethical AI principles. Key components of ProFinda’s proprietary architecture include:
- Proprietary Ontology and Graph Database: ProFinda’s intellectual property is centered on a proprietary global ontology combined with localized client-specific ontologies. This graph database powers a multidimensional relationship model, allowing for a much richer representation of relationships between skills, job titles, companies, and qualifications, processing information far beyond a simple hierarchical taxonomy.
- Global and Local Ontology Power: The Global Ontology is vast and dynamic, mapping over 90,000 skills, 900,000 terms, and 15 million interconnects, continuously updated from job boards, market data, and expert curation. This provides comprehensive, externally validated skill suggestions, ensuring a broad skill landscape is considered. The Local Ontology integrates an organization’s unique internal skills, job titles, and qualifications, reflecting its unique nuances. This combination leads to
superior matching accuracy and reduced bias.
2. High-Risk AI Applications and Risk Control
Under the EU AI Act, ProFinda has identified several product applications classified as high-risk, each of which has specific, built-in risk mitigation measures.
AI Application
Risk for Bias
Data Value / Risk Relevance
Total Risk Assessment
Generative AI Role Descriptions
Medium
Low
Low
Profile Match
Low
High
Low
Skills Suggestions
Low
High
Low
CV Parser
Low
High
Low
Skills Framework
Low
Medium
Low
Bias Control at the Core of Matching (Profile Match & Smart Search)
ProFinda is designed from the ground up to prevent and mitigate the bias inherent in data and models.
- Data Isolation (PII Exclusion): The Profile Match AI module is deliberately separated and explicitly does not synchronize or use Personally Identifiable Information (PII) during match calculations. This ensures matching decisions are not generated based in any personal data.
- Deterministic and Predictable Results: The AI module is built separately from the API and the main database. It does not learn from one search to the next; each match request starts from zero and applies the same defined rules. This design prevents ProFinda from biasing a search by identifying or prioritizing the most frequently selected candidates.
- Matching Scores and Transparency: The Profile Matching algorithm calculates separate scores for Direct skills, Related skills, and Indirect skills. The final score combines these factors , and all scores are surfaced separately in the interface to provide a clear and transparent understanding of the final result.
Human Oversight and Mitigation
For all AI-generated content and suggestions, ProFinda maintains a strict policy of human intervention:
- Generative AI Role Descriptions: This application uses external Large Language Models (LLMs), which may or may not have inherent bias in their proprietary training data. Guardrails are in place to evaluate the generated descriptions. The output is treated as a suggestion that is evaluated by the user and requires manual selection.
- Skills Suggestions: This feature is a key way ProFinda mitigates bias in self-reported skills: Our machine learning identifies similar profiles and suggests skills that will enhance the user’s profile, regardless of PII factors such as gender or ethnicity. Crucially, no skill is automatically added; it requires a human decision and manual intervention to ultimately add the skill.
CV Parser
The CV parser uses internal algorithms and NLP to identify and extract relevant entities like job titles, skills, experience, and education. This high-value content is critical for matches and recommendations. The low risk for bias is maintained by ensuring that the NLP services (Aida and Hal) are in constant evolution, adapting to the changes in language and skills.
3. Governance and Continuous Improvement
ProFinda is committed to following current and future legal requirements, including the EU AI Act, and maintaining best practices in data quality and risk mitigation.
Risk Mitigation Processes
- Technical Documentation and Record Keeping: Comprehensive technical documentation ensures all aspects of the EU AI Act and conformity assessment are applied. Automated processes ensure full traceability, and any events are safely stored.
- Training Data Practices: All data undergoes data governance and management practices to ensure it is relevant and free from errors. A critical safety measure is the exclusion of PII from training data.
- Continuous Evaluation and Metrics: ProFinda uses an iterative process of testing and evaluating improvements. This includes employing rigorous
bias metrics (e.g., Statistical Parity Diff, Disparate Impact) to measure differences across groups and continuously test models against diverse datasets. - Human Oversight and Training: As a way of ensuring risks are minimized and user rights are protected, Human Oversight is a key practice. Training is provided to engineering and product teams to help them understand and mitigate unconscious bias when coding and designing system outputs.
4. Future Vision: Enabling Collaborative Automation
ProFinda is dedicated to continuous innovation, focusing on providing clients with even greater control and automation capabilities.
- Model Context Protocol (MCP): ProFinda is actively working to wrap all its powerful services within a Model Context Protocol (MCP). This initiative will expose APIs in a secured, standardized, accessible manner, making them readily consumable by external systems and agents while ensuring governance is in place.
- Intelligent Automation through Agents: ProFinda is building an internal multi-agent system to increase its capabilities and allow it to perform every action that headless ProFinda offers. This opens opportunities for building automation and optimization on day-to-day operations.
- Your Bedrock for Innovation: ProFinda aims to be the “bedrock” upon which organizations can build amazing, tailored automations. We provide the complex AI and scalable infrastructure, freeing client teams to innovate without the overhead of maintaining the underlying AI technology.
