Mohammadjavad Mehditabar

Machine Learning Engineer with 4 years of experience across academia and industry, specializing in applied R&D, data science, and the performance optimization of LLMs and Agentic AI. My work spans building end-to-end ML pipelines, benchmarking LLMs on cost-latency-accuracy trade-offs, and profiling hardware-level compute bottlenecks. I focus on aligning models and developing trade-off-aware agentic workflows to build highly efficient, scalable AI systems.

Technical Skills

Programming

Python C# Javascript SQL

Frameworks & Tools

PyTorch vLLM LangChain Transformers Pandas Numpy Docker Git

Research & Strategy

Qualitative Research Survey Design Customer Discovery Business Modeling AI Commercialization

Professional Experience

Graduate Research Assistant

SMART Lab, Dalhousie University (Nova Scotia, Canada) - Advisor: Tushar Sharma

Jan 2025 – Present
  • LLMs Performance Benchmarking Code: Developed two novel, robust, and fair Multi-Criteria Decision Making (MCDM) methods with less than 5% performance variation under severe noise conditions. Evaluated 22 LLMs on performance and throughput trade-offs on the LiveCodeBench and CodeXGLUE benchmarks. Investigated advanced serving techniques (continuous batching, speculative decoding) to maximize GPU utilization.
  • Code Inefficiencies Detection Paper: Created a taxonomy of inefficiency patterns in code, curated a high-quality dataset of 3,000 code pairs, and developed an agentic pipeline using a frontier LLM with dynamic context retrieval achieving 94% accuracy.
  • Meta-Data Aware Movie Recommender Code: Devised an end-to-end modified GPT-based recommender system fusing textual metadata with user interaction sequences via multi-task learning and LoRA fine-tuning. Improved genre-awareness F1-score by 37% and boosted overall recommendation accuracy (NDCG) by up to 5%.

R&D Analyst

Lab2Market Canada

Sep 2025 – Dec 2025
  • Conducted R&D and 80+ user interviews to evaluate technical solutions optimizing the cost-accuracy-latency trade-off in AI-based products.
  • Validated 20 architectural hypotheses with stakeholders ranging from CXOs to developers, shaping novel value propositions for highly efficient and privacy-preserving Agentic AI systems handling sensitive data.

Research Assistant

Iran University of Science Technology (Tehran, Iran) - Advisor: Sauleh Eetemadi

Jul 2021 – Jun 2024
  • Consensus on Machine Learning Models: Improved the predictive accuracy of traditional ML and ensemble models by developing a novel, interactive consensus approach, outperforming standard baselines.
  • Long Corpus Personality Detection Code: Proposed a novel approach to classify long documents by leveraging two consecutive BERT models (Hierarchical BERT) to address transformer context limitations.

Data Scientist

Dadmatech (Tehran, Iran)

Jun 2022 – Jan 2023
  • Engineered automated ETL pipelines for Twitter, combining web crawling with regex-based text structuring. Developed a custom Django application for user data collection and management.
  • Built and optimized a comprehensive suite of machine learning solutions, progressing from classical algorithms to advanced neural networks and transfer learning.

Selected Publications

Watts This Smell: A Comprehensive Taxonomy of Software Energy Smells
Mehditabar, M., Rajput, S., & Sharma, T.
IEEE International Conference on Software Maintenance and Evolution (ICSME, CORE A), IEEE, 2026. Early acceptance rate: 6%.
Paper
BRACE: Unified Benchmarking of Accuracy and Energy for Code Language Models
Mehditabar, M., Rajput, S., Mastropaolo, A., & Sharma, T.
Journal Ahead Workshop at ICSE 2026.
Paper
A consensus-based approach to improve the accuracy of machine learning models
Karamdel, H., Ashtiani, M., Mehditabar, M. J., & Bakhshi, F.
Evolutionary Intelligence, 1–22, 2024.
Paper
MBTI Personality Prediction Approach on Persian Twitter
Fatehi, S., Anvarian, Z., Madani, Y., Mehditabar, M., & Eetemadi, S.
WiNLP at EMNLP Workshop, 2022.
Paper Poster

Notable Projects

Personality Detection (NLP)

Source Code

  • Introduced novel model architectures for classification and incorporated non-text features. Implemented sentence generation with GPT and automated crawling/TF-IDF pipelines.

Visual Question Answering

Source Code

  • Built a custom transformer-based MiniVQA model from scratch. Extracted image embeddings via ResNet, and question embeddings using Distil-BERT.

Covid Detection on X-Ray Chest Images

Source Code

  • Employed data augmentation and transfer learning via SqueezeNet. Fine-tuned the last two layers to classify COVID presence with rigorous sensitivity/specificity testing.

Education & Teaching

Dalhousie University

Master of Computer Science (GPA: 4.3/4.3)

Jan 2025 – Present
  • Supervisor: Prof. Tushar Sharma

Iran University of Science and Technology (IUST)

B.Sc. Computer Engineering (GPA: 3.94/4)

Sep 2019 – Sep 2023
  • Supervisor: Prof. Sauleh Eetemadi

Teaching Experience

Software Engineering (TA) Computational Intelligence (TA) Algorithm Design (TA) Theory of Automata (TA) Data Structures (Mentor) Advanced Programming (Mentor)

Honors & Awards

Winner, Lab2Market Industrial Proposal ($10,000)

2025
  • Won an award for proposing a novel approach to reduce the cost of integrating generative AI for small and medium-sized enterprises.

Research Scholarship Grant ($43,000)

2025
  • Awarded for M.Sc. Program at Dalhousie University.

Academic Excellence

2019 - 2023
  • Ranked 3rd out of 90 students in B.Sc. program.
  • Ranked in the top 0.4% of 144,000 participants in the National University Entrance Exam.