AI Engineer

What I worked on

CSIRO's Data61 is Australia's national science agency for data and AI. I joined the digital forensics group to work on explainable deepfake detection — the kind of system that needs to defend its conclusions to a human investigator, not just output a probability.

The system

A multimodal detection pipeline:

• Detection — CNN + Vision Transformer ensemble across frames
• Explainability — SHAP and LIME for saliency heatmaps showing which regions drove the prediction
• Captioning — CLIP + attention-based captioning describing suspicious regions in natural language
• Contextual explanation — Llama 3.1 (via Ollama) with custom prompt engineering, producing forensic narratives that an investigator can read
• Deployment — wrapped as REST APIs so the model integrates into existing forensic workflows without rewriting the front-end

Outcomes

• Published as DF-P2E at ACM Multimedia 2025 — multimodal, explainable, interactive deepfake detection for non-experts
• Live Hugging Face Space for interactive exploration
• Research poster presented at the DFRWS Digital Forensics Conference
• Production REST API deployed into the team's analysis workflow