Open to opportunities

deepu d

@deepud

Message

Machine learning research intern building adversarially robust and optimized LLM/vision systems.

India

Message

What I'm looking for

I’m looking for a role where I can build and evaluate robust ML systems and optimize LLM/vision inference end-to-end, with hands-on engineering in PyTorch and CUDA, clear benchmarks, and room to grow into impactful research-to-production work.

I’m a machine learning research intern focused on making models both resilient and performant in real-world conditions. At VISTORA AI, I engineered adversarial robustness pipelines (FGSM, PGD, and C&W) on ResNet-50 and ViT, reaching a 92% attack success rate with imperceptible perturbations.

I also built practical safety tooling—developing a backdoor detection framework with Neural Cleanse and spectral signature analysis (87% of poisoned samples in CIFAR-10, <3% false positive rate). I implemented certified defenses like randomized smoothing and improved robustness by 23% on the AutoAttack benchmark while maintaining 89% clean accuracy.

Beyond research, I ship end-to-end systems: DENTEX applies Mask R-CNN (ResNet-101-FPN) to dental X-rays with 0.78 mAP@IoU=0.5 and optimizes inference to 180ms per image. I’m also working on production-scale LLM inference optimization using vLLM and SGLang (PagedAttention KV cache management, continuous batching, FP8/GPTQ) with benchmark-driven throughput and latency improvements.

Experience

Work history, roles, and key accomplishments

ML Research Intern

Vistora AI

May 2025 - Jul 2025 (2 months)

Engineered adversarial robustness pipelines using FGSM, PGD (ϵ = 0.03), and C&W on ResNet-50 and ViT, achieving 92% attack success rate with imperceptible perturbations. Built a Neural Cleanse backdoor detection framework (87% poisoned-sample detection at <3% false positives) and improved robustness by 23% on AutoAttack while maintaining 89% clean accuracy via randomized smoothing and adversarial

Carlini & Wagner (C&W) Attacks Backdoor Detection (Neural Cleanse)AutoAttack Benchmarking PyTorch ResNet And Vision Transformers