This text describes a complex recommendation system designed to connect people needing expertise with qualified experts. Unlike typical recommendation systems, this one faces unique challenges like the limited capacity of experts and the high cost of bad recommendations. It employs a hybrid retrieval method using Reciprocal Rank Fusion to combine results from multiple independent expert finders. The scoring mechanism is a weighted combination of various factors, including explicit directional fit, semantic similarity, skill overlap, capacity, expert quality, specialty match, and fairness.Expert quality uses a saturating exponential function to account for experience, preventing veterans from completely dominating. Fairness incorporates a logarithmic decay for exposure, ensuring less frequently recommended experts still get opportunities. Importantly, the system operates as a global assignment problem rather than individual top-N recommendations to prevent the most popular experts from being overwhelmed. A greedy allocation algorithm is used, prioritizing higher-scoring pairings while respecting expert capacity limits.The data layer focuses on defensible statistics to determine what skills are valuable. This involves source weighting with exponential decay for recency, weighted medians to handle skewed data distributions, and stratification by role, seniority, and experience to avoid conflating skill value with seniority. Bootstrap confidence intervals are used instead of point estimates for skill value, providing a more robust measure.A critical dry run revealed significant flaws, including an overzealous exclusion rule that severely limited expert availability and a tendency for recommendations to lack clear justifications. The trending badge also exhibited inflated values due to a lack of a meaningful baseline. These findings highlight the importance of careful implementation and continuous evaluation in complex recommendation systems.