EdgeTal: On-Device Agentic RAG for Privacy-Preserving Talent Discovery

Every AI-powered recruitment tool on the market today shares one architectural assumption: candidate data must travel to the cloud to be useful. CVs, salary expectations, performance signals, and career histories — some of the most sensitive personal data an organization holds — are routinely transmitted to third-party inference APIs, creating regulatory exposure under GDPR-class regimes and a data-sovereignty problem no consent checkbox actually solves.

EdgeTal inverts that assumption. It is an agentic Retrieval-Augmented Generation system that runs entirely on-device: a local vector store over candidate and role data, a quantized small language model for generation, and an agentic orchestration layer that plans multi-step talent-discovery queries — matching, screening, and shortlist reasoning — without a single byte of candidate data leaving the device. No cloud inference, no third-party API dependency, no data residency ambiguity.

The research contribution is demonstrating that agentic RAG — typically assumed to require frontier-scale cloud models — is viable within edge constraints (memory, thermal, and latency budgets of commodity hardware) for a real, high-stakes enterprise workload. The system achieves a Mean Average Precision (mAP) of 0.88 in retrieval matching and an average end-to-end query latency of 3.2 seconds on commodity M-series hardware using a quantized 8B-parameter local model.

EdgeTal sits within Knovik's broader privacy-preserving AI research program alongside SemanticGuard (with CDAC, La Trobe University), forming a two-pronged thesis: where data cannot be protected in transit, don't transmit it at all.