Senior ML / Research Engineer · Ph.D.

LLMs · Multimodal AI · Computer Vision · Machine Unlearning · Efficient Training

Senior ML engineer and published researcher working on large language models, multimodal AI, computer vision, machine unlearning, and efficient training systems. I focus on taking research-grade ideas all the way into production at scale.

• Senior Machine Learning Engineer at TikTok — novel MLLM architectures for Trust & Safety, shipped to production.
• Research Fellow at AIML, University of Adelaide and Visiting Research Scientist at CSIRO — co-investigator on an A$1.2M grant training frontier-scale foundation models on 256× NVIDIA H200 GPUs.
• Co-founder of A2.AI — an applied-AI venture.
• Previously shipped ML research into VFX pipelines on Mad Max: Furiosa, Mortal Kombat II, Deadpool, Mickey 17, Sonic 3, Sinners, Michael, and A Complete Unknown at Rising Sun Pictures.

Read the longer story, publications, and interactive explainers at arpit2412.github.io.

Full list on Google Scholar.

• US Provisional Patent — Attention mechanism for compute- and memory-efficient LLM training (filed 2026).
• UK Design Patent (granted, No. 6520933) — AI-Assisted Rural & Indigenous Healthcare Robot.
• ICML 2025 Best Reviewer — Gold Award.
• Invited Speaker, MLSS Melbourne 2026.

• Co-investigator on an A$1.2M ResetData grant to train frontier-scale foundation models (language, multimodal, reasoning) on a 256× NVIDIA H200 cluster.
• Own end-to-end training methodology, alignment and controllability research, and stability/throughput validation of the multi-million-dollar datacenter.
• Authored compute- and memory-efficient LLM training that cuts wall-clock time and peak GPU memory simultaneously — covered by a US provisional patent.
• Research on machine unlearning, LoRA / PEFT, and stable vision-language alignment — accepted at CVPR 2026 (SineProject), multiple NeurIPS 2026 submissions, TPAMI under review.
• Joint appointment at CSIRO advising on responsible-AI and trustworthy LLM / MLLM systems.

Designing and shipping novel multimodal LLM (MLLM) architectures for Trust & Safety — production models that reason over image, video, and text together at platform scale.

• Shipped MLLM architectures lifting business-data AUC by +2–3%, with a further +5% from ensembling and distillation.
• Own the full production loop — retraining, evaluation, and deployment of safety models — with cross-functional engineering and product teams.
• Mentor engineers, drive research → engineering handoffs, and co-author top-tier peer-reviewed publications through sustained academic partnerships.

Vision + language stack

Architectures & techniques I build on — modern multimodal designs that inform the Trust & Safety models:

• LLaVA-OneVision — a unified large multimodal model spanning single-image, multi-image, and video understanding.
• MoVA (NeurIPS 2024) — a mixture of vision experts that adaptively routes and fuses task-specific encoders via context-aware routing, since no single encoder wins on every image type.

Machine-learning research shipped into production VFX pipelines at Rising Sun Pictures — deepfake, gaze-estimation, and generative shot work. Credits on IMDb.

Furiosa: A Mad Max Saga 2024 · VFX ML Research	Mortal Kombat II 2025 · VFX ML Research	Deadpool & Wolverine 2024 · VFX ML Research
Mickey 17 2025 · VFX ML Research	Sonic the Hedgehog 3 2024 · VFX ML Research	Sinners 2025 · VFX ML Research
Michael 2026 · VFX ML Research	A Complete Unknown 2024 · VFX ML Research	La Brea 2023 · VFX ML Research

• InstanceGM — Instance-dependent noisy-label learning via graphical modelling (WACV 2023).
• PASS-NoisyLabel — Peer-agreement based sample selection (IMAVIS 2025).
• NoiseRateLearning — Graphical-model-based noise-rate estimation.
• Generative-Adversarial-Network — A tour of major GAN families.
• arpit2412.github.io — Portfolio and interactive ML explainers.