.
I build frontier AI systems and teach them to forget.
PhD, machine learning. Research fellow at AIML and CSIRO. Senior ML engineer at TikTok. Co-founder of A2.AI.
Defence labs to frontier datacenters.
Eight years across research, industry, film, and national-scale AI infrastructure.
May 2025 - Present
Research Fellow & Visiting Research Scientist
AIML, University of Adelaide · CSIRO
A$1.2M Grant Co-Investigator
- Co-investigator on an A$1.2M ResetData grant to train frontier-scale foundation models (language, multimodal, reasoning) on a 256× NVIDIA H200 cluster.
- Owned end-to-end training methodology, alignment, controllability research, and stability/throughput validation of the multi-million-dollar datacenter.
- Authored compute and memory-efficient LLM training that simultaneously reduces wall-clock time and peak GPU memory; covered by a US provisional patent.
Oct 2024 - Present
Senior Machine Learning Engineer
TikTok
Trust & Safety Research
- Designed and shipped novel MLLM architectures for Trust & Safety: +2-3% AUC on business data, +5% additional lift via ensemble and distillation.
- Stack: SigLIP, CLIP, SAM, Co-DETR, DINOv2, ConvNeXt vision backbones paired with Phi, Gemma, Mistral LLMs.
- Owned retraining, evaluation, and deployment of production safety models with cross-functional engineering and product teams.
Jun 2023 - Sep 2024
Machine Learning Developer (Research)
Rising Sun Pictures
VFX ML Research
- Designed a novel background-augmentation and occlusion-aware loss for deepfake pipelines, reducing FID by 15%.
- Integrated into shipping VFX workflows on Mad Max: Furiosa, Deadpool, Mickey 17, La Brea, Sonic 3, and Sinners.
- Shipped a production gaze-estimation model improving facial authenticity in VFX shots, reducing gaze error by 4%.
Jan 2020 - Jun 2023
Machine Learning Researcher
Adelaide Business School & UoA
Applied CV & NLP for Market Intelligence
- Built CV/NLP market-intelligence systems covering 5,000+ companies; +12% sentiment accuracy and +7% data-driven decision quality.
- Designed an NLP expert-recommendation system with custom embeddings scaling to 200,000+ professional profiles.
- Architected an automated DL framework for security-patch classification across 10,000+ vulnerabilities.
May 2018 - Mar 2019
Applied ML Research Intern
DRDO & WESEE (Indian Navy)
Defence R&D
- Built CV / DIP algorithms for satellite-imagery analysis at >1M-image scale (DRDO).
- Delivered 20+ mission-critical algorithms improving real-time decision speed by 21% and operational efficiency by 25% for naval weapons systems (WESEE).
Research-grade ideas, shipped at production scale.
Senior ML engineer and published researcher specialising in large language models, multimodal AI, computer vision, machine unlearning, and efficient training systems. Co-founder of A2.AI (a2ai.com.au).
Recognised for shipping research-grade systems into production at scale, from TikTok's Trust & Safety MLLMs to VFX pipelines on Mad Max: Furiosa, Mortal Kombat II, Deadpool, Mickey 17, Sonic 3, Sinners, Michael and A Complete Unknown.
Inventor on a US provisional patent (attention mechanism) and a granted UK design patent (AI-Assisted Rural & Indigenous Healthcare Robot). Co-investigator on an A$1.2M grant powering frontier-scale training on 256× NVIDIA H200 GPUs.
Education
Ph.D., LLMs / MLLMs, Generative AI & Computer Vision
Australian Institute for Machine Learning, University of Adelaide
Nov 2021 - Jan 2025
M.S., Artificial Intelligence & Data Science
The University of Adelaide
Jul 2019 - Jun 2021
B.Eng., Computer Engineering
Rajasthan Technical University
Aug 2015 - Jun 2019
Pretrain. Fine-tune. Deploy. Unlearn.
I have owned every stage of the model lifecycle, including the one most people never reach: teaching a deployed model to forget.
Pretrain
Frontier models from scratch
Data pipelines, tokenizers, and distributed training runs on a 256× NVIDIA H200 cluster. Stability, throughput, and training methodology owned end to end.
A$1.2M ResetData grant
Fine-tune
Alignment and PEFT
LoRA with learnable rank, regularizers that stop fine-tuning from forgetting, and vision-language alignment that stays stable under pressure.
NeurIPS 2026 submissions · US patent filed
Deploy
Production at platform scale
Trust & Safety MLLMs serving TikTok, with the full retraining, evaluation, and deployment loop. ML research shipped into nine feature-film VFX pipelines.
TikTok · Rising Sun Pictures
Unlearn
Making models forget
The step most teams never get to: provable, bounded machine unlearning, so deployed models can remove what they should no longer know.
SineProject · CVPR 2026
Research
Making models efficient, multimodal, and able to forget.
Six threads run through my work, from machine unlearning to million-token attention.
Machine Unlearning
CVPR 2026 · NeurIPS 2026 (sub.)Making models forget, provably and stably. SineProject (CVPR 2026) and bounded parameter-efficient unlearning in LLMs.
Efficient Training & PEFT
US Patent (filed) · NeurIPS 2026 (sub.)Compute- and memory-efficient LLM training (US provisional patent), learnable-rank LoRA, and regularizers against LoRA forgetting.
Multimodal LLMs
CVPR 2026 · TikTok productionVision-language alignment and production MLLM architectures that reason over image, video, and text together.
Generative Models
NeurIPS 2026 (sub.) · 9 filmsDiffusion guidance without retraining (STRIDE), GANs and deepfake pipelines for film VFX, NeRF and Gaussian Splatting.
Long Context & Attention
Attention Atlas · vLLM deep-diveAttention mechanisms from first principles, from the original Transformer to linear attention and million-token contexts.
Robust Learning
ECCV 2024 · WACV 2023 · TPAMI (sub.)Instance-dependent noisy-label learning via graphical models and peer-agreement sample selection.
Selected publications
SineProject: Machine Unlearning for Stable Vision-Language Alignment
A. Garg, H. Saratchandran, S. Lucey
LR-LoRA: Parameter-Efficient Fine-Tuning with Learnable Rank
A. Garg, S. Lucey, H. Saratchandran
Mask the Target: A Plug-and-Play Regularizer Against LoRA Forgetting
R. Xu*, A. Garg* (co-first), H. Saratchandran, S. Lucey
From datacenters to cinema screens.
Frontier training runs, production safety models, film VFX, and a venture of my own.
AIML · ResetData Grant
Frontier-Scale Foundation Model Training
Co-investigator on an A$1.2M grant training language, multimodal, and reasoning models on a 256× NVIDIA H200 cluster, owning training methodology, alignment research, and stability/throughput validation of the datacenter.
Production ML
Trust & Safety MLLMs at TikTok
Novel multimodal LLM architectures reasoning over image, video, and text at platform scale, +2-3% AUC on business data, +5% further via ensembling and distillation. Owned retraining, evaluation, and deployment.
Nine films carrying shipped ML research
Interactive deep-dives into how AI actually works.
Visual, hands-on explainers, from attention mechanisms to inference engines.
Recognition along the way.
ICML 2025 Best Reviewer: Gold Award
Recognised among the top reviewers worldwide.
Investigator, A$1.2M ResetData Grant
Training large foundation models on 256× NVIDIA H200 GPUs.
Invited Speaker, MLSS Melbourne 2026
By invitation from Maincode.
Open-Source Impact
140,000+ visits across public ML repositories.
US Provisional
Attention Mechanism for Neural Networks
Compute & memory-efficient training of large language models. · Filed May 2026
UK Design (Granted)
AI-Assisted Rural & Indigenous Healthcare Robot
Class 24, Medical Equipment · UK Intellectual Property Office · No. 6520933 · 29 April 2026
Latest
Invited speaker: MLSS Melbourne 2026. Lecturing at the Machine Learning Summer School, Melbourne (by invitation from Maincode).
UK Design Patent granted (No. 6520933). AI-Assisted Rural & Indigenous Healthcare Robot, Class 24, Medical Equipment. UK Intellectual Property Office.
SineProject accepted at CVPR 2026. Machine unlearning for stable vision-language alignment. First-author work with Saratchandran and Lucey.
US Provisional Patent filed. Attention mechanism for neural networks, compute and memory-efficient LLM training, productionised in internal pipelines.
AEON submitted to TPAMI. Adaptive estimation of instance-dependent ID/OOD label noise for robust learning. arXiv:2501.13389.
PASS published in Image and Vision Computing. Peer-agreement based sample selection for training with instance-dependent noisy labels.
ICML 2025 Best Reviewer: Gold Award. Recognised among the top reviewers worldwide.
A$1.2M ResetData grant: investigator. Lead investigator for frontier-scale foundation model training on 256× NVIDIA H200 GPUs.
Contact
Building something at the frontier?
Let's talk.
Research collaborations, advisory work, speaking, or ambitious products. I'm always open to a good conversation about hard AI problems.