Dilawar Shah

About

Computer Vision & Generative AI Intern and undergraduate researcher building production-grade ML systems, including controlled synthetic data generation with Stable Diffusion/ControlNet and diffusion fine-tuning with LoRA, plus research projects spanning RAG pipelines, CUDA-accelerated neural network training, and MLOps ETL workflows on AWS/Redshift/Airflow. Skilled in PyTorch, Transformers, LoRA/knowledge distillation, inference optimization, and GPU performance tuning.

Experience

• Computer Vision & Generative AI Intern · NexPred Solutions10-01-2025 – 12-31-2025

  • Architected and deployed controlled synthetic image generation pipeline using Stable Diffusion + ControlNet, enabling production-ready training dataset for vehicle damage detection models • Fine-tuned diffusion models with LoRA (8–12M trainable parameters) for domain adaptation, improving visual realism and downstream model accuracy by 25–30% • Optimized inference latency from 9.5s → 6.5s per image (30% reduction) through DDIM scheduling and kernel optimization, reducing cloud compute costs by ~35% at scale • Generated 3,000+ high-resolution synthetic images (512×512 / 768×768) meeting production quality standards; validated with downstream computer vision models Tech Stack: PyTorch, Diffusers, LoRA, CUDA, AWS GPU instances

• Undergraduate Researcher – Final Year Project · FAST National University08-01-2025 – 04-30-2026

  • Developed TraCC, a production-grade multi-label code comment classification system; achieved F1 = 0.671 and competition score of 0.761 at NLBSE 2026 (peer-reviewed, ACM-affiliated workshop) • Applied knowledge distillation (CodeBERT → TinyBERT) reducing model size by 88% (125M → 14.5M parameters) and inference latency by 33%, enabling real-time deployment in resource-constrained environments • Mitigated severe class imbalance using LLM-based data augmentation (Google Gemini API), generating 4,700 synthetic training samples and improving minority-class F1 scores by 23–63% • Published peer-reviewed paper accepted at NLBSE 2026; demonstrated advanced techniques in model compression and imbalanced learning Tech Stack: Hugging Face, Transformers, PyTorch, Google Gemini API, imbalanced-learn

Education

• FAST National University of Computer and Emerging SciencesB.S., Computer Science

Similar talent on Pangea

Panav J.Machine Learning Intern | Deep Learning • NLP • MLOps (PyTorch/FastAPI/Docker)Delhi, IndiaAymen F.Junior AI Engineer | Deep Learning & Computer Vision | Full-Stack (FastAPI/React)Chott Meriam, TunisiaSamarth Z.Founding Engineer & Research Engineer | Backend Systems, Distributed Data Pipelines & LLM/ML Fine-TuningLos Angeles, CACamila A.Senior Machine Learning Engineer | Applied Scientist (Computer Vision, MLOps, GenAI) | 6+ YearsCampinas - State of São Paulo, Brazil