I’m an undergraduate researcher building ML systems that are rigorous enough to trust and practical enough to ship: clear objectives, reproducible pipelines, and evaluation that matches real constraints (compute, latency, data quality, and downstream requirements). I work across generative modelling, computer vision, NLP, and measurement-heavy applied ML.

• Generative Systems & Control: Score-based diffusion (SDEs/ODEs), controllable generation (classifier-free/gradient guidance), inference-time latent optimisation, and flow matching.
• Edge Deployment & Efficiency: Post-training quantisation (PTQ/QAT), structured pruning, knowledge distillation, system-aware neural architecture search, and latency-constrained inference (TFLite/ONNX).
• ML Systems Engineering: Reproducible experimentation (tracking/versioning), CI/CD for stochastic pipelines, artefact lineage, and scalable data loading.
• Robustness & Reliability: Out-of-distribution (OOD) detection, uncertainty quantification (conformal prediction), evaluation under deployment constraints, and failure-mode analysis.
• AI for Healthcare & Impact: Causal inference, fairness-constrained optimisation, interpretability in high-stakes decision making, and privacy-preserving data pipelines.
• Computer Vision: Representation learning (SSL), fine-grained visual recognition, and geometric deep learning.
• Natural Language Processing: Parameter-efficient fine-tuning (LoRA/adapters), retrieval-augmented generation (RAG), and processing noisy/low-resource text.
• Mathematical Structure: Optimisation landscapes in deep learning, stability analysis of generative flows, and generalisation bounds.

• Centre for Constrained Image Generation (Columbus, OH) — Post-Generation Team (Oct 2025 – Present): constraint-aware diffusion/inpainting pipelines; evaluation harnesses; reliability and export-validity gates; team + sprint execution.
• Science of Sadharan × Prakash Labs (Stanford) — Research Lead, M&E Vertical (Jul 2024 – Oct 2025): NLP + CV pipelines for impact evaluation; scaled data-collection automation; reporting and analysis workflows.
• KCDH-IITB (Koita Centre for Digital Health) — Research Intern (Oct 2023 – Jul 2024): mental-health NLP pipelines and agent-based simulation for intervention evaluation.

• Preprint: Inference-Time Loss-Guided Colour Preservation in Diffusion Sampling
  Training-free diffusion control using inference-time perceptual objectives and region stability mechanisms.
  arXiv: https://arxiv.org/abs/2601.17259

• Constraint-aware generation pipelines (diffusion/inpainting)
  Built and evaluated 6 end-to-end architectures; selected a pipeline that reliably outputs 600 DPI / 300 DPI with compliance checks and export-validity gates.

• Deepfake Detection on Edge (ResNet-50 + transfer learning + quantization)
  Edge-deployable detector with 82% accuracy on the quantized model, tracking latency/size trade-offs for constrained devices.

• Vision Transformer on CIFAR-10 (from scratch)
  75% test accuracy (+12% over CNN baseline), 22% lower training time (gradient checkpointing), and 3x faster convergence (LR warmup + cosine decay).

• Content-Aware Pagination (notes → printable PDFs)
  Content-aware splitting of long note images into clean printable PDFs with user-controlled page formats.
  Repo: https://github.com/AngadSingh22/Content-Aware-Pagination