Francesco Cavalli

Objectives. This submission examines how an integrated investigative AI platform supports law-enforcement units tasked with sexual-crimes and CSAM imagery investigations. It addresses a recurring operational gap: standalone detection tools verify individual files, but units run cases composed of many files across long timeframes, multiple jurisdictions, and several analysts. The objective is to show how combining detection layers with structured case management and a convergence engine changes both the speed and the defensibility of
investigative outcomes.

Methods. The platform integrates several detection layers as first-class signals within a shared case workspace: deepfake analysis across image, video, and audio modalities, distinguishing real, fully synthetic, and hybrid content; face matching 1:N against controlled victim reference libraries and offender watchlists; voice matching 1:N across an internal reference pool, including cross-domain pools shared between directorates under appropriate governance; context intelligence on audio, surfacing dialect, recurring names, location references, and behavioural patterns; geolocation analysis from visual cues; and file-level forensics establishing source-device characteristics, editing history, and production attribution. Case management enforces chain of custody, role-based access, model versioning, and audit trail. A
convergence engine proposes cross-evidence connections automatically through a case graph.

Results. The integrated approach compresses investigation timelines at every stage. Triage of mixed real, synthetic, and hybrid material completes in a single pass, surfacing priority files for victim identification within hours rather than weeks. Face and voice matching run automatically against reference libraries as evidence is ingested, without analyst-initiated comparison. Context intelligence reduces hundreds of hours of audio to reviewable patterns in days. Cross-layer corroboration emerges as a routine output rather than weeks of manual cross-referencing. The platform consistently moves investigations from referral to first operational action in a fraction of conventional timeframes.

Conclusions. Detection alone produces verdicts on files; integration produces case outcomes. Key learnings. Hybrid content implies a real victim. Cross-layer corroboration is the strongest evidentiary foundation. Reproducibility and auditability determine whether findings survive prosecutorial and judicial scrutiny.