CV
An overview of research, experience, publications, and background.
Contact Information
| Name | Meghdad Kurmanji |
| Professional Title | AI Research Scientist |
| mk2296@cam.ac.uk | |
| Phone | +44 7949 726 826 |
| Location | Cambridge, UK |
| Website | https://meghdadk.github.io |
Professional Summary
AI Scientist with 8+ years of experience designing and deploying scalable generative AI systems across research and production settings. Expertise spans the full lifecycle of machine learning systems, from prototyping novel models to building robust, production-grade pipelines for enterprise applications. Proven ability in delivering impactful solutions in large-scale model training, multimodal learning, and AI safety, with publications in top-tier venues including NeurIPS and ICLR and competitive research funding for privacy-preserving and safety-critical AI systems.
Experience
IQVIA
AI Research Scientist
London, UK
2026 – present
- Designed and trained transformer-based foundation models on longitudinal clinical data, enabling structured reasoning over long patient timelines.
- Developed agentic LLM systems for automated knowledge extraction and verification from large document collections using LangGraph and retrieval pipelines.
University of Cambridge
Postdoctoral Research Associate
Cambridge, UK
2024 – 2026
- Secured a GBP 150k Foresight AI Safety grant to advance research on mechanistic machine unlearning and AI safety.
- Co-led a EUR 530k-funded research project on scalable decentralized LLM pre-training with privacy-preserving and robust training pipelines.
- Developed novel machine unlearning techniques for efficient and secure data removal, addressing regulatory requirements such as GDPR compliance.
- Designed and evaluated distributed LLM training systems using PyTorch, Transformers, and MosaicML in large-scale environments.
- Built scalable evaluation pipelines for LLM fine-tuning across downstream tasks, improving reproducibility and benchmarking.
- Published research in top-tier venues including NeurIPS and ICLR.
University of Warwick
Graduate Research Assistant
Coventry, UK
2020 – 2024
- Pioneered machine unlearning algorithms addressing data deletion and privacy requirements in large-scale ML systems.
- Designed a continual learning framework achieving more than 10x throughput improvement compared to prior approaches.
- Developed a machine unlearning algorithm outperforming prior state-of-the-art methods by up to 10 percent across benchmarks.
- Secured GBP 150k research funding from Huawei for machine learning-based database indexing systems.
- Led collaborations with Google DeepMind, including co-organizing the NeurIPS 2023 Machine Unlearning Challenge.
- Co-authored more than 7 publications in NeurIPS, SIGMOD, and CIDR.
Iran Telecommunication Research Center (ITRC)
Data Engineer
Tehran, Iran
2019 – 2020
- Built an end-to-end big data pipeline from crawling to Hadoop and OLAP, reducing data onboarding time by 5x.
- Implemented scalable ETL workflows, achieving a 5x query speed-up.
- Integrated Elasticsearch with PowerBI dashboards, reducing reporting delays by 60 percent.
Refah Retail Chain Stores Co.
Machine Learning Engineer
Tehran, Iran
2017 – 2019
- Led development and deployment of a real-time computer vision system for customer footfall analysis, achieving 81 percent accuracy across 20 stores.
- Developed in-store heatmap analytics to optimize staffing and store layout decisions.
- Built a multimodal recommendation system combining LSTM and CNN models, increasing customer engagement by 15 percent.
- Implemented time-series regression models for customer behavior prediction.
Sensifai
Deep Learning Engineer
Belgium (Remote)
2016 – 2017
- Improved production audio classification accuracy by 9 percent using multimodal transfer learning techniques.
- Built an 88 percent accurate music mood classifier using spectrogram-based CNN models.
- Optimized distributed video crawling pipelines, achieving 1.8x throughput.
Education
University of Warwick
PhD in Computer Science
Coventry, UK
2020 – 2024
- Thesis: Adaptability of ML-Based Database Systems (SIGMOD Honorable Mention Award).
- Conducted the first study of data deletion in learned database systems (SIGMOD 2024).
- Developed SCRUB, a state-of-the-art unlearning algorithm for large-scale deep models (NeurIPS 2023).
- Proposed DDUp, a framework for efficient data insertion in learned database systems (SIGMOD 2023).
- Collaborated with Google DeepMind to launch the NeurIPS Machine Unlearning Challenge.
Tarbiat Modares University
MSc in Computer Science
Tehran, Iran
2014 – 2017
- Dissertation: Hand Gesture Recognition Using 2D and 3D Convolutional Neural Networks from Video.
- GPA: 3.67/4.
Isfahan University of Technology
BSc in Computer Engineering
Isfahan, Iran
2010 – 2014
- GPA: 3.65/4.
Skills
Core Expertise
LLM pre-training and fine-tuning, Decentralized ML, Machine Unlearning, AI Safety
Programming
Python, C++, SQL, Bash
ML Frameworks
PyTorch, TensorFlow, Hugging Face, Transformers
Distributed Systems
PyTorch DDP and FSDP, Ray, MosaicML, Slurm
MLOps and Cloud
Docker, CI/CD, AWS, Azure, Databricks, MLflow, Weights and Biases
Data Systems
SQL, NoSQL, OLAP, Hadoop, Learned Indices
Awards
Secured GBP 150k Foresight AI Safety Grant
2026
NeurIPS Top Reviewer Recognition
2025
SIGMOD Jim Gray Doctoral Dissertation Honorable Mention
2025
EUR 530k SPRIN-D Grant
2024-2025
Co-lead, CambridgeFlower
Organizer, NeurIPS Machine Unlearning Workshop
2023
Best Presentation Award, WPCCS, University of Warwick
2021
Graduate Scholarship
2020-2024
GBP 25k per year, University of Warwick
Research Grant
2020-2024
GBP 15k per year, Huawei
Publications
DEPT: Decoupled Embeddings for Pre-training Language Models
ICLR
2025
Top 1 percent paper on pre-training language models with decoupled embeddings.
Bridge the Gaps between Machine Unlearning and AI Regulation
NeurIPS
2025