Hi, I am

Anshu.

I am an Applied AI Research Engineer at GovTech Singapore, Singapore’s main government technology agency, where I build and evaluate AI systems across generative modelling, synthetic data, and data privacy.

My work spans AI research, full-stack engineering, and product development, allowing me to take projects from early research and rapid prototyping through user validation, production deployment, and rigorous evaluation.

I work closely with global experts across industry and academia and remain deeply curious about emerging advances in both communities. I enjoy identifying promising ideas, testing them in real-world settings, and shaping them into reliable AI products.

Previously, I was a Research Assistant at the National University of Singapore (NUS), working with Prof. Mohan Kankanhalli and Dr. Shaojing Fan on computer vision research under constrained visual settings. I also hold a Master's in AI from NUS (dissertation track).

Blogs

Understanding Inference Economics with DeepSeek V4

Watercolor illustration of Anshu at CVPR

Image and Video Generation at CVPR 2026 (in-person)

More here Medium.

Selected work

SynthSite synthetic video generation and evaluation pipeline

Video diffusion · Wan2.2-I2V-A14B · Qwen2.5-VL curation · LoRA · 227 rare scenarios · VBench · open-vocabulary detection

SynthSite — Synthetic video generation & benchmarking

SynthSite began as a practical deployment problem — evaluating safety site-analytics vendors with almost no footage of rare hazards — and grew into research. It is a hybrid local–cloud video generation pipeline with human curation and safety annotation, LoRA fine-tuned to synthesize rare worker-under-suspended-load scenarios, with agency domain experts reviewing the outputs. Published at the CVPR 2026 SynData4CV workshop; dataset and code on Hugging Face.

PyTorch · VAE / GAN / diffusion · model routing & cascades · HPO · AWS

Mirage — Whole-of-Government synthetic data platform

Mirage is the government's self-service platform for mock and synthetic data generation, used in production by 30+ agency teams. I led the build-vs-buy evaluation against vendor systems, then built the generation stack in-house — PyTorch modeling across copula, VAE, GAN, and diffusion generators, adaptive model routing and cascading, automated hyperparameter optimization, and multi-dimensional privacy–utility evaluation.

Explainable classification · RAG + vector DB · RAG evaluation · HCI-informed design · multimodal document AI · agentic orchestration · human-in-the-loop

DataSharingAssist

I ideated and product-led DataSharingAssist, an AI assistant for explainable data classification and data-sharing guidance — recommendations and justifications grounded in government policies through retrieval, with the full reasoning process visible to the user. It gained traction with 20+ agencies through in-person pilots with data teams and policymakers, was presented at the IEEE Symposium on Privacy Expectations in New York, and was selected into the government-wide Build for Good hackathon for its high impact.

Experience

2022–now.
Applied AI Research Engineer, Government Technology Agency (GovTech), Singapore.
- Lead R&D in synthetic data generation, including an accepted CVPR 2026 SynData4CV paper (paper; dataset and code) and applied solutions for challenging agency use cases.
- Co-led the design and productisation of Mirage, a whole-of-government synthetic data generation platform.
- Owned and served as product lead for an LLM/RAG-powered decision-support system for data classification.
- Authored a synthetic data primer, reviewed and well received by leading academic, industry, and regulatory experts.
2021–2022.
Research Assistant (Computer Vision & NLP), NUS Centre for Research in Privacy Technologies.
- First authored “Human Attributes Prediction under Privacy-Preserving Conditions.” ACM Multimedia 2021. [Oral]
- Built an end-to-end geocoded social-media intelligence platform.
2019–2021. Master of Computing (Artificial Intelligence), National University of Singapore — dissertation in computer vision.
2016–2019. Software Engineer, DBS Bank and HashedIn Technologies (Deloitte) — customer-facing web products from ideation and piloting through production.

Selected Research Outputs

2026. Anshu Singh^*, Alejandro Seif^*. "Privacy-Aware Synthetic Video Benchmarking and Relational Evaluation for Worker-Under-Suspended-Load Detection." 3rd Workshop on Synthetic Data for Computer Vision (SynData4CV), CVPR 2026. [*equal contribution]
Multimodal large language models · hybrid open/closed-model pipeline · LoRA fine-tuning · open-vocabulary detection and segmentation · spatiotemporal reasoning · video quality metrics · human annotation
2025. Authored a primer on synthetic data — a playbook covering generation methods, evaluation, risks, and practical guidance for researchers, practitioners, and decision makers — reviewed by Prof. Mohan Kankanhalli, Prof. Xiaokui Xiao, Dr. Shlomi Hod, Dr. Yihao Ang, Prof. Anthony Tung, and Singapore's PDPC.
LLM-driven data synthesis · generative AI methods · multi-dimensional evaluation · risks and challenges · practical synthesis pipelines
2023. Authored Sharing Data with Differential Privacy: A Primer and the accompanying benchmarking repository, reviewed by Prof. Xiaokui Xiao and Dr. Damien Desfontaines.
Open-source DP tooling · distributed data processing · scalable analytics · cloud deployment · performance engineering
2021. Anshu Singh, Shaojing Fan, Mohan Kankanhalli. "Human Attributes Prediction under Privacy-Preserving Conditions." ACM Multimedia 2021. [Oral]
Custom deep learning architecture · pose-guided feature fusion · human-centered AI · dataset curation · explainable AI