Multimodal deepfake detection
Matched real/fake evaluation protocol (AUC / FPR on honest negatives), a 13-paper audit narrowed to 5 audio-visual models, and fine-tuning on Malaysian demographic data.
Nine years training, compressing and deploying deep learning systems across speech, vision and multi-agent LLM pipelines, from synthetic data to TensorRT on a Jetson to a weekly production run nobody has to babysit.
A podcast episode drops, and a manager agent fans the transcript out to parallel LLM workers, gates every edit with deterministic scripts, and ships bilingual Malay + English subtitles. It writes its own retrospective when it's done.
Read the architecture →
Matched real/fake evaluation protocol (AUC / FPR on honest negatives), a 13-paper audit narrowed to 5 audio-visual models, and fine-tuning on Malaysian demographic data.
LangGraph multi-agent system with plan decomposition and a self-healing data-analyst agent, plus an LLM-as-judge eval service and a containerised MCP server for live financial data.
Photorealistic Unity/Blender data pipeline with domain randomisation for rare edge cases, QAT + TensorRT for edge budgets, and multi-camera detections projected into a live warehouse digital twin.
Custom YOLO detectors in an always-on AWS pipeline over live RTMP streams, then ported to Jetson Nano/Xavier with TensorRT + DeepStream for real-time FPS.
Patch-based training and inference over >1B-pixel drone maps that kept global spatial coherence, with a GPS lat/lon for every detected palm tree.
Regional challenge entry focused on spatial reasoning. Also top 5 at the Seedance “Build with AI” hackathon 2026 (AI video generation).
Let's build
something.
Research roles, consulting, or a model that has to work outside the notebook. Send a note and I'll reply by email.