BLR23:50:36
···18:20:36
00:00
X0.0000
Y0.0000
← Back to projects
Client Delivery • Prismforce

Graph-Based Skill Recommendation Engine

+30% relevance, sub-50ms inference. Weighted directed graph over multi-level skill hierarchies - dynamic node updates without batch retraining, deterministic traversal under production load.

+30% Recommendation relevance
<50ms Inference latency (99th percentile)
0 Batch retraining on taxonomy expansion

The Problem

The recommendation system treated related skills as independent nodes with no structural modeling - hierarchical relationships were ignored entirely. Every taxonomy expansion triggered full batch retraining, blocking updates until recompute completed. Inference latency under production concurrency exceeded the sub-50ms SLA required for live platform use.

The System

Real-time recommendation engine using a weighted directed graph encoding multi-level skill hierarchy relationships as typed edges with dynamic weight updates. Lightweight mathematical scoring heuristics minimize computational overhead per inference call - no neural inference at query time. The update model handles dynamic node additions without full graph recomputation: adding a new skill inserts a node and its edges; existing traversal paths are unaffected.

Deterministic traversal logic produces consistent outputs under frequent profile and taxonomy updates. The graph structure constrains traversal depth, so worst-case latency is knowable before deployment. Latency profiled at the 99th percentile under realistic production concurrency on NVIDIA T4 - not average latency under ideal conditions.

Outcomes

+30% improvement in recommendation relevance from structural hierarchy modeling. Sub-50ms inference on a single NVIDIA T4 under production load - 99th percentile SLA met. Dynamic taxonomy expansion no longer triggers batch retraining. Consistent outputs maintained under frequent profile and taxonomy changes.