Talenthacked: a Client-side Evidence-grounded Ai Stack for Uk Global Talent Visa Adjudication
Year: 2025
Type: article
Abstract: We present TalentAct, a zero-install, browser-native AI stack that couples (i) hybrid BM25 + dense retrieval with Maximal Marginal Relevance, (ii) per-claim NLI verification, and (iii) temperature-scaled calibration (ECE = 0.036) for UK Global Talent Visa (GTV) adjudication. Running entirely client-side (TensorFlow.js + WASM), the system outputs citation-anchored evidence maps in < 3.8 s on a £180 Android device. A 500-case prospective pilot across three endorsing bodies improved Cohen’s κ from 0.70 → 0.88, cut median review time 43% and reduced inter-body refusal-rate variance by 31%. A counterfactual regression using Home-Office micro-data implies national deployment would retain ≈ 385 additional tech founders yr⁻¹, unlocking £1.38 bn (95% CI 1.05–1.71) extra venture investment. All code, WASM bundles and 1k synthetic adversarial cases are released under MIT licence.
Institutions: University of Bradford, University of Hull
Language: English
Topic: AI and HR Technologies
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7081943958
FWCI:
Cites:
Cited by:
Related to: