System Design

Decoupled Design

TechWolf implements a decoupled design, which enables end-to-end explainability, strong control over bias in training, and more efficient development cycles. Inferring skills from documents, aggregating documents for an employee, and creating recommendations are trained and evaluated separately.

For each data type supported by our product, we train individual NLP models. Our skill timeline mechanism aggregates the identified skills across all data of an employee into a complete and up-to-date skill profile. Skill-based recommendations operate on the anonymous skill profile level and take into account adjacencies between skills in TechWolf’s ontology. Isolating the different parts in this system allows us to clearly define the scope of each component, and validate or improve their correct functioning efficiently.

Explainability

TechWolf is end-to-end explainable, both for skills and recommendations. For skills inferred for an employee, we mainly provide explanations on which documents or events impacted the inference, but more detailed explanations that show the impact of sections of each document can be provided too. For recommendations, a skill-level breakdown of contributing factors is provided as an explanation.

Thanks to our decoupled architecture and transparent models, explanations from our product reflect what’s actually happening, rather than being created after the fact.

Configurability

Data landscapes can be vastly different between organisations or departments, with differences in data availability, volume, and quality. The separation between inference on a document level, and the combination of those data points into an employee skill profile, allows us to configure and optimise the setup of our product for maximum accuracy in each scenario. TechWolf can control for noise in the data and ensure balanced skill profiles even if input data is skewed or inconsistent.

As a part of our core product functionality, customers can import and govern their taxonomy and job framework inside the system, ensuring the AI works in line with existing structures in their ecosystem.