Amir Mahdi

Between data and decisions lies a space where systems learn to think.
This is a record of building, studying, and shaping that space through research and real-world experimentation.

Portfolio

A curated overview of work shaped by inquiry and experimentation spanning research projects, technical achievements, and selected credentials.

Work Experience

2025 – Present

AI Research Engineer

Deep Computer Vision Co.

Developing state-of-the-art AI systems and machine learning models, with a strong focus on computer vision. Actively conducting research and contributing to academic publications.

2024 – 2025

Data Science Intern

Dr. Samati’s Lab, Sharif University of Technology

Collaborated on research in Agentic AI, working closely with academic teams to analyze data and develop intelligent agent-based models.

2024

Machine Learning Bootcamp

University of Kashan

Completed a one-month intensive bootcamp focused on machine learning concepts, data preprocessing, model training, and evaluation using Python and relevant libraries.

Education

2021 - 2025

Bachelor of Engineering - BE, Computer Engineering

University of Guilan

Sep 2018 - Jun 2021

High School Diploma

Alborz Highschool

Certificates

Machine Learning Specialization

Coursera - 2023

Docker

Arjang institute - 2024

Data Analytics Bootcamp

Tehran University - 2022

Projects

Ideas take shape here, through systems that learn, adapt, and act.
Each project reflects a process: from problem framing to methodological choices and final implementation.
Technical notes, source code, and supplementary materials are shared where applicable.

Persian Sentiment Analysis

This project introduces a custom Transformer-based architecture designed for sentiment analysis on Persian text reviews from the SnappFood dataset. By leveraging a pretrained BERT tokenizer for text encoding and fine-tuning a lightweight Transformer encoder, the model effectively classifies reviews as positive, neutral, or negative. This approach addresses challenges posed by Persian morphology and script, demonstrating the effectiveness of custom architectures for low-resource NLP tasks.

PyTorch Docker Django

View on GitHub

Image Captioning

Automatically generating descriptive captions for images is a challenging task with applications in accessibility, image retrieval, and content generation. This project focuses on building a deep learning model using the Flickr8k dataset to generate accurate captions for images by combining Convolutional Neural Networks (CNNs) for image feature extraction and Recurrent Neural Networks (RNNs) for language generation.

PyTorch OpenCV

View on GitHub

Age Estimation with ResNet

This project aims to improve the accuracy and adaptability of age estimation models by leveraging the power of ResNet architectures. The current age estimation models, primarily based on PyTorch and convolutional layers, often face limitations in terms of accuracy and generalization across diverse datasets like UTKFace.

OpenCV PyTorch

View on GitHub

Language Modeling with LSTM and AWD-LSTM on WikiText-2

Language modeling is a critical task in natural language processing (NLP) that involves predicting the next word in a sequence given its preceding context. This project builds and evaluates two models for this task: Base Model: A traditional LSTM-based language model. Improved Model: An AWD-LSTM model that incorporates advanced regularization techniques like weight dropping, variational dropout, and optimization strategies such as ASGD to enhance performance and generalization.

Pytorch

View on GitHub

Competitions & Achievements

Challenges tackled, skills demonstrated, and recognition earned through competitive programming and AI competitions.

Kaggle - CMI Detect Behavior with Sensor Data

🏆 Ongoing

2025 - Predicting Body Focused Repetitive Behaviors from a Wrist-Worn Device

Kaggle - MITSUI&CO. Commodity Prediction Challenge

🏆 Ongoing

2025 - Developing a robust model for accurate and stable prediction of commodity prices

Publications

Research at the intersection of learning, language, and autonomy.
Featured works include peer-reviewed papers, ongoing collaborations, and contributions to the broader AI research community.

DreamerAgent: A Computational Model of the Lacanian Unconscious Using Memory-Driven Signifier Activation and Neurochemical Emotional Modeling

2025 Preprint avaiable on SSRN

AmirMahdi Aramideh, Tina Falak Doust Noshary, Ehsan Kozegar

This paper presents DreamerAgent, a novel AI framework that integrates Lacanian psychoanalytic theory with modern language and vision models to simulate a symbolic unconscious. Using dream narratives, emotional modeling, and memory-driven signifier activation, it enables AI agents to exhibit human-like symbolic complexity, contradiction, and emergent psychological behaviors.

Paper Code