Case study · AI engineering
Improving expert matching with semantic search
Cognivox Labs designed and productionised a semantic matching system for a German B2B platform, replacing the limitations of an existing keyword-driven approach with context-aware expert matching and integrating the new capability into the client’s established Laravel application.
~76% to ~98%
Match relevance in internal benchmark testing
+22 percentage points
improvement in match relevanceInternal benchmark comparison using the same representative matching scenarios and ranked-result relevance assessment.
Overview
Enhancing an existing matching workflow
The client operates a B2B platform that connects user requirements with relevant experts.
An existing matching mechanism was already in place, but its reliance on keyword similarity limited the quality of results when requirements became more detailed or when equivalent expertise was expressed using different terminology.
Cognivox Labs was engaged to develop a more context-aware approach to matching, validate its performance against the existing system and, once proven, integrate it into the client’s deployed application.
Challenge
Improving relevance beyond keyword similarity
The existing workflow relied primarily on keyword-based matching between user requirements and expert information.
This approach performed reasonably for direct terminology matches, but became less reliable when the language used in a requirement differed from the terminology used in an expert profile.
The engineering challenge was therefore not simply to retrieve profiles containing similar words. The system needed to identify experts whose experience was relevant to the underlying meaning and intent of the requirement.
At the same time, any new matching capability had to fit into the client’s existing Laravel and PostgreSQL platform without requiring a wider rebuild of the production application.
Solution
A dedicated semantic matching service
Cognivox Labs developed the matching capability as an independent Python service rather than embedding the new logic directly into the Laravel application.
The matching pipeline uses semantic representations of user requirements and expert information to retrieve candidates based on contextual similarity. Candidate results are then further evaluated and ranked to improve the relevance of the final recommendations.
This architecture allowed the matching logic to be developed and iterated independently while preserving a clear boundary between the AI capability and the client’s existing product.
Separating the matching engine from the core application was an important architectural decision. Laravel remained responsible for the product workflow and application state, while the Python service handled the semantic matching process through a defined API interface.
- Client platform
Laravel application
- Matching request
New requirement passed to matching service
- Semantic matching
Python / FastAPI
- Semantic retrieval
- Candidate evaluation
- Ranking
- Results
Ranked expert matches returned
to the existing platform
Validation
Benchmarking the new matching approach
Before production integration, the semantic matching system was evaluated against the existing keyword-based workflow using the same representative matching scenarios.
The assessment focused on the relevance of the ranked expert results returned for each requirement.
Internal benchmark testing showed match relevance increasing from approximately 76% with the existing approach to approximately 98% with the semantic matching system.
Difference: +22 percentage points
Internal benchmark comparison using consistent matching scenarios and ranked-result relevance assessment.Production integration
Integrating the matching engine into the live platform
Following validation, the second phase of the project focused on production integration.
The client’s application was already deployed using Laravel with PostgreSQL. Rather than introducing the matching logic into the existing PHP codebase, Cognivox Labs exposed the Python implementation through a FastAPI service.
Laravel communicates with the service through an API boundary, while background jobs handle matching asynchronously so that the process does not block the user-facing application.
When a new requirement is created, the application can trigger the matching workflow automatically, receive the ranked expert results and persist those results within the existing platform.
- Requirement created
- Laravel background job
- Semantic matching service
- Ranked results
- Stored in platform
The integration also retained manual controls for testing, re-running matches and operational administration.
Outcome
Improved matching within the existing product architecture
The project delivered a measurable improvement in matching relevance while allowing the client to retain its existing application architecture and product workflows.
The semantic matching capability was moved from an independently validated algorithm into an operational production service integrated directly with the platform’s requirement-creation workflow.
Match quality
Internal benchmark match relevance increased from approximately 76% to 98%.
Product integration
Semantic matching became part of the existing platform workflow rather than remaining a standalone tool or prototype.
Architecture
The AI matching service remained independently maintainable while the core Laravel application continued to manage the product workflow.
Engineering
Technology used
- Laravel
- PostgreSQL
- Python
- FastAPI
- Semantic search
- Vector retrieval
- Background processing
Building an AI capability into an existing product?
Cognivox Labs designs and integrates production AI systems around existing software, data and operational workflows.
Discuss a projectClient identity withheld under confidentiality agreement.
← Back to case studies