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

Introduction: Every time a recruiter uploads a resume to a cloud-based hiring platform, the candidate's personal data leaves their control. AI-powered screening has made hiring faster, but by routing sensitive information through external servers, it trades off against modern privacy law. Most mobile recruitment tools also still rely on exact keyword matching rather than semantic understanding.

Objectives: This study asks whether a smartphone can perform intelligent candidate search and analysis entirely on device, without any cloud connection.

Methods: EdgeTal was developed as a native Android application built around a three-stage on-device Agentic RAG pipeline. Recruiter queries are encoded as semantic embeddings, matched against candidate profiles through vector similarity search, and interpreted by a quantised Gemma-2B (Int4) language model running entirely on the device. The system was evaluated on 1,500 resumes across two devices, a Google Pixel 7 Pro and a Xiaomi Redmi Note 7.

Results: Semantic search responded in under 100ms on the Pixel 7 Pro and under 300ms on the Redmi Note 7. For abstract and role-based queries, keyword matching returned no relevant results, whereas semantic retrieval consistently surfaced relevant candidates.

Conclusions: These results show that privacy-preserving, intelligent recruitment can run on existing mobile hardware rather than dedicated cloud infrastructure, given an appropriate on-device architecture.

Keywords