Rohan Manzoor

Experience

• 

  National University of Sciences and Technology (NUST)

  Sept 2022 - Nov 2023

  • Identified biases impacting disabled individuals in 15 prominent vision datasets through advanced tools like ViTs and Yolo. Implemented MiDas and SAM for free space estimation. • Achieved precise detection of objects with as few as 700 instances, by using Selenium for web scraping from platforms like TripAdvisor and Airbnb along with StyleGAN 2-ADA, ControlNet, and LORA fine-tuned diffusion. Utilized Grounding Dino for efficient annotation along with ensemble models for further improvement in accuracy.• Achieved successful black box adversarial attacks on SIFT and SURF through the strategic application ofRayS, SignOPT, and HopSkipJump. Collaborated on the development of an in-house attack method• Used several tools including python, pytorch, tensorflow, keras, pandas and others for data analysis and data visualization Show less Served as TA for Dr. Wajahat's Machine Learning and Deep Learning course. Aided in content delivery and student education, encompassing sharing forefront research, publications, and instructing model design and deployment.• Created lectures on cutting-edge topics: self-supervised learning, diffusion, transformers, and more.• Conducted PyTorch labs, sessions, and quizzes for learning and assessment, offering constructive feedback and support.• Guided, assessed, and upheld standards for student projects, exams, and vivas to ensure academic integrity.• Mentored students in projects, reports, and presentations, offering feedback, suggestions, and the latest research insights Show less

  • Research Assistant

    Jun 2022 - Nov 2023

  • Graduate Teaching Assistant

    Sept 2022 - Jun 2023
• 

  AxcelerateAI

  Jun 2023 - Jun 2024

  Machine Learning Engineer

  • Developed backend using LangChain, GPT-4o, Gemini, Groq, Ollama to automate weekly payment reset-tlement by dynamically analyzing financial statements and MCA contracts.• Integrated advanced RAG techniques (re-ranker) and advanced prompting(chain of thought, tree ofthought) for one-click generation of new settlement scores and justifications.• Improved stable diffusion inference and memory utilization by 27% and 35% via various optimizations• Developed automatic masking for image inpainting, enhancing restoration efficiency by 50% on human eval• Improved image fidelity by finetuing diffusion models using LoRA, dreambooth and textual inversion• Improved accuracy by 50% and inference speed by 200% using CLIP based zero shot image classifier• Developed end-to-end data and model training pipelines, seamlessly integrating data from multiple sources• Fine-tuned Whisper for Arabic ASR on custom dataset for call center audio transcription Show less

• 

  NETSOL Technologies Inc.

  Jul 2024 - now

  Machine Learning Engineer

  • Fine-tuned Gemma 2b with LoRA on synthetic data for enhanced search functionality.• Developed agentic workflows for database question answering (text-to-SQL) with self-reflection, RAG, and CoT.• Built an automated tool to complete RFPs utilizing Pinecone, MongoDB, and finetuned SLM.

• 

  Information Technology University

  Sept 2024 - Jan 2025

  Graduate Teaching Assistant

  Working as a graduate teaching assistant for Generative AI course offered by Dr Mohsen Ali