Projects

Robust Offline Spike Sorter (ROSS)

2020
Associated with PhD Thesis, University of Tehran

Matlab spike sorting extracellular recordings

Matlab spike sorting, extracellular recordings, neural data analysis, signal processing, clustering, offline analysis, machine learning Neural activity monitoring is one of the bases of understanding the brain behavior. The recorded extracellular potential is a combination of multiple neurons activities corrupted by noise. A main step in analyzing the extracellular data is to differentiate among different neuron activities. Spike sorting is the process of assigning each detected spike to the corresponding neurons. ROSS is a MATLAB based offline spike sorter software which helps the researchers to do automatic and manual spike sorting tasks efficiently. Currently it provides t-distribution and skew-t based methods for automatic spike sorting. Several functions are considered for modifying the auto-sorting results, such as merging and denoising. Also, useful visualizations are provided to get better results.

Deep Learning for MRI Reconstruction

2025
Associated with LiteMRI

Python MRI reconstruction inverse problems

Python MRI reconstruction, inverse problems, reconstruction pipelines, robust evaluation, reproducible workflows, medical imaging, deep learning, inference Development and evaluation of deep learning methods for MRI reconstruction as part of the machine learning team behind a clinically used software product.The work covers reconstruction pipelines based on Recurrent VarNet, DIRCN, and image-to-image models. My role spans training, refining, and evaluating models across varied experimental conditions, running systematic optimisation experiments, and contributing to reproducible workflows for the machine learning model lifecycle.The central challenge is recovering clinically useful images from incomplete or degraded measurements, which places a strong emphasis on quantitative evaluation and failure analysis rather than qualitative inspection of outputs alone. This connects directly to broader research interests in inverse problems, learned representations, consistency constraints, and robustness across acquisition conditions.

Regex Engine & grep-Style Search (Rust)

2025
Associated with CodeCrafters

Rust regex engine pattern matching

Llama Chatbot in Rust (llama.cpp)

2024
Associated with Personal Project

Rust llama.cpp local LLMs

CNN Training & Inference in Rust

2024
Associated with Personal Project

Rust deep learning CNN

K-Means Clustering in Rust

2024
Associated with Personal Project

Rust clustering

Spatial Frequency Representation in the Inferior Temporal Cortex

2025
Associated with Brain Computing Lab, University of Tehran

Matlab spatial frequency neural decoding

Matlab spatial frequency, neural decoding, IT cortex, population coding, electrophysiology, statistical analysis, machine learning My doctoral research examined how spatial frequency shapes object representation in the primate inferior temporal (IT) cortex, the final stage of the ventral visual pathway. Using neural population recordings from macaque monkeys viewing spatial-frequency-filtered objects, the analysis linked spatial frequency content to category coding across levels of abstraction.The main finding is that spatial frequency information is explicitly represented in IT cortex, with low spatial frequencies decoded earlier than high spatial frequencies. Follow-up work showed that mid-level category boundaries remain robust across spatial frequency bands, while superordinate coding depends on high spatial frequency content and fine subordinate discrimination requires the full spectrum.This repository contains the replication code for the decoding analyses under both fast and slow visual presentations, covering population decoding, statistical analysis, and the figures reported in the paper.

Brain-Inspired Spatial-Frequency-Aware Networks

2023
Associated with Brain Computing Lab, University of Tehran

Python spatial-frequency-aware networks neuroscience-inspired ML

Python spatial-frequency-aware networks, neuroscience-inspired ML, feedback connections, frequency filtering, image classification, deep learning, CNN A translation of neurophysiological findings on spatial frequency coding into artificial network design. Having found that the primate visual system processes coarse and fine spatial structure on different timescales, this project asked whether an artificial classifier benefits from the same organisation.The resulting architecture uses two streams that separately process low and high spatial frequency content, together with a feedback mechanism inspired by the recurrent connectivity of the biological visual system. Coarse structure provides rapid context while fine detail supports precise discrimination.The approach improves classification performance under spatial frequency filtering on CIFAR-100, indicating that biologically motivated frequency-aware processing can increase robustness when input detail is degraded.

Cage Camera for Rodent Behavioural Monitoring

2021
Associated with University of Tehran

Python rodent monitoring Raspberry Pi

Python rodent monitoring, Raspberry Pi, Flask, OpenCV, behavioral neuroscience, database, computer vision, inference An open-source remote video monitoring system for rodents housed in Faraday cages, built to support behavioural neuroscience experiments.The system runs on Raspberry Pi cameras with a Flask server and a web interface for recording, live viewing, and camera parameter control. Recordings and session metadata are stored in a database, and the services are managed through systemd so that acquisition continues reliably over long experimental sessions without manual intervention.The project addresses the practical side of behavioural data collection: recording control, laboratory constraints, usable interfaces for experimenters, and integration of software with existing experimental workflows.

Mousi — Mouse Tracking and Behavioural Analysis

2021
Associated with University of Tehran

Python mouse tracking PyQt GUI

Python mouse tracking, PyQt GUI, OpenCV tracking, behavioral analysis, trajectory extraction, computer vision, inference A desktop application for tracking mice in behavioural video recordings, built with Python and PyQt around an OpenCV-based tracking pipeline.The tool provides a graphical interface for loading recordings, configuring tracking parameters, monitoring the extracted trajectory, and exporting results for downstream behavioural analysis. It was designed for use by experimenters rather than developers, so most of the effort went into making the tracking configurable and its output inspectable.Together with the cage camera system, this project connects computer vision directly to behavioural neuroscience applications, from acquisition through to quantitative movement analysis.

Video Content Creation and Curation

2019
Associated with Adak AI

Python video analytics object recognition

Python video analytics, object recognition, object tracking, scene classification, automated tagging, genre classification, recommendation systems, computer vision, deep learning, inference This project is a comprehensive video analysis system that includes several components: Object recognition: The system is capable of recognizing objects within a video and tracking their movements. Scene classification: The system can classify the different scenes within a video, such as indoor or outdoor, night or day, etc. Automated tagging: The system can automatically tag videos with relevant information, such as the objects present, the scene, and any other relevant details. Genre classification: The system can classify videos into different genres, such as action, drama, comedy, etc. Actor age, gender, and emotion: The system can detect the age, gender, and emotion of actors in a video. Recommender system: The system can recommend videos for further extension based on the above analysis and viewing history.This project is associated with Viddsee, a video entertainment platform based in Singapore. By incorporating all these components, the project aims to provide a more personalized and engaging video viewing experience for users.This project brings several advantages for the company, Viddsee: Improved user engagement: By providing personalized video recommendations and in-depth video analysis, the system can help improve user engagement and increase user satisfaction. Better video classification and organization: The automated tagging, genre classification, and scene classification capabilities of the system can help Viddsee better organize and classify their video content, making it easier for users to find the videos they are interested in. Enhanced video advertising: Viddsee can target advertisements more effectively, resulting in increased revenue from advertising. Data-driven insights: The system generates a wealth of data about the video content and the user behavior, which Viddsee can use to gain valuable insights into their business and make data-driven decisions. Overall, this project can help Viddsee enhance the user experience, improve their video content classification and organization, increase revenue from advertising, and gain valuable insights into their business.

Movie Genre Detection

2019
Associated with Communication and Computation Lab, University of Tehran

Python movie genre detection multimodal learning

Multimodal Book Genre Prediction

2025
Associated with University of Tehran

Python multimodal learning book genre prediction

Python multimodal learning, book genre prediction, VisionMamba, ViT, vision-language models, feature fusion, image classification, deep learning, inference A multimodal system that predicts book genres from cover images by combining complementary visual encoders with text representations.The architecture pairs ViT and VisionMamba image encoders, which capture global and local visual structure respectively, with text embeddings derived from vision-language-model-generated descriptions of each cover. Rather than relying on OCR to read the limited text printed on a cover, the VLM descriptions supply richer thematic and narrative cues that the raw tokens miss.Evaluated on 57,000 covers spanning 30 genres, the method reaches 63.31% Top-1 and 83.03% Top-3 accuracy. The project provided experience with transformer and state-space visual backbones, multimodal feature fusion, and evaluation on a substantial dataset.

Human Pose Estimation and Motion Capture

2018
Associated with Adak AI

Python C++ human pose estimation motion capture

Python, C++ human pose estimation, motion capture, OpenPose, keypoint detection, 3D pose estimation, computer vision, deep learning, inference A group of related projects on extracting body posture and movement from video without wearable sensors.The work covered 2D human pose estimation in unconstrained scenes, facial keypoint detection in video with a comparative evaluation of competing algorithms, multi-person body motion capture, monocular 3D pose estimation from a single camera, and 3D pose estimation from an eight-camera setup. One strand mapped estimated 2D poses onto a Blender rig to drive character animation, and the monocular system was developed towards a commercial version aimed at education.These projects built practical experience with detecting and representing keypoints, posture, and movement from image sequences, which underpins behavioural AI systems that combine kinematic information with learned visual representations.

Real-Time Face Search in Video

2018
Associated with Adak AI

Python C++ face search in video face recognition

Python, C++ face search in video, face recognition, real-time inference, CUDA, video retrieval, computer vision, deep learning A deep learning system for locating a target face across video content, combining face recognition with efficient inference.The pipeline reaches approximately 60 frames per second on Full HD video using a single GTX 1080 Ti, which required attention to recognition quality and computational performance together rather than in isolation. A companion project targeted the opposite end of the hardware spectrum, running face recognition on a Raspberry Pi with an explicit time-versus-quality trade-off.The problem is closer to query-by-example retrieval than to frame-level classification, since it involves finding semantically matching instances across an entire video.

Retrieval-Augmented Generation for LLM Systems

2024
Associated with Adak AI

Python retrieval-augmented generation LLMs

Python retrieval-augmented generation, LLMs, knowledge graphs, multi-hop question answering, re-ranking, chain-of-thought, NLP, inference Collaborative research on grounding large language model outputs in retrieved evidence, spanning two complementary directions.The hybrid retrieval approach combines sentence-window and parent-child retrieval with a re-ranking stage, improving both the relevance of retrieved passages and the faithfulness of generated answers to their source material. The knowledge-graph extension (KG-RAG) targets multi-hop question answering by decomposing questions and using in-context learning with chain-of-thought prompting, so that multi-step retrieval produces explicit and inspectable reasoning chains instead of opaque single-shot answers. The method requires no training, which keeps it adaptable across different knowledge graphs with minimal human effort, and improves accuracy on multi-hop questions in the MetaQA benchmark.Both projects treat retrieval as infrastructure for inspectable reasoning, addressing when a language model should be trusted and when structured evidence should constrain it.

Persian Speech Recognition with Whisper

2025
Associated with Adak AI

Python Persian speech recognition Whisper fine-tuning

Python Persian speech recognition, Whisper fine-tuning, automatic speech recognition, WER/CER metrics, dataset pipeline, audio processing, deep learning, inference An end-to-end pipeline for Persian automatic speech recognition, built around fine-tuning and evaluating OpenAI Whisper large-v3-turbo.The project unifies several public and in-house Persian speech corpora into a single dataset stored in Parquet format, then fine-tunes and evaluates the model with word and character error rate metrics. Training runs are supported both locally and on Modal cloud infrastructure, with dependencies managed through uv for reproducibility.This builds on earlier speech work including a Persian speech-to-text system based on Mozilla DeepSpeech and a lightweight recogniser optimised for Raspberry Pi 4 deployment, where model compression was necessary to meet edge constraints.

Fine-Tuning LLaMA 3.2-Vision for Product Descriptions

2025
Associated with Personal Project

Python multimodal LLM LLaMA 3.2 Vision

Python multimodal LLM, LLaMA 3.2 Vision, LoRA fine-tuning, 4-bit quantization, vision-language training, text generation, deep learning, inference Fine-tuning a multimodal large language model to generate short, SEO-friendly product descriptions directly from product photographs.The project adapts LLaMA 3.2-11B-Vision with LoRA on a dataset of kid’s clothing images, using 4-bit quantization and the Unsloth training stack so that the whole workflow fits on a single consumer GPU. It covers dataset preparation, parameter-efficient fine-tuning, and inference.The result is a practical demonstration of pulling structured, usable text out of messy visual input under realistic hardware constraints.

Model Optimization: Quantization and Pruning

2024
Associated with Personal Project

Python model compression quantization-aware training

Python model compression, quantization-aware training, pruning, ResNet, efficient inference, optimization, deep learning, CNN An implementation and comparison of model compression techniques for ResNet architectures in PyTorch, covering post-training quantization, pruning, and quantization-aware training (QAT).The project examines the practical trade-offs between model size, inference cost, and predictive performance, with QAT recovering accuracy that naive post-training quantization gives up. It complements earlier edge-computing work on emotion recognition and face recognition deployed to resource-constrained hardware.Efficiency work of this kind matters wherever deployment constraints sit alongside representation quality, from mobile inference to training and serving large video models.

Weak Supervision with Snorkel

2023
Associated with Personal Project

Python weak supervision Snorkel

Python weak supervision, Snorkel, labeling functions, probabilistic labels, image classification, data-centric AI, deep learning An image classification study built on weak supervision rather than manually labelled data, using Snorkel on the Open Images dataset.Instead of hand-annotating examples, the project defines labelling functions that encode noisy heuristics, then combines their conflicting votes into probabilistic training labels which are used to train a downstream PyTorch classifier.The approach is relevant wherever ground-truth annotation is expensive or requires expert time, which is common in biomedical and scientific imaging settings.

Diffusion Models from Scratch

2024
Associated with Personal Project

Python diffusion models DDPM

Python diffusion models, DDPM, U-Net, generative modeling, image generation, training from scratch, deep learning, CNN A denoising diffusion probabilistic model (DDPM) implemented from first principles in PyTorch, trained to generate anime-style faces.The repository builds the full pipeline without relying on a diffusion library: the forward noising schedule, a U-Net noise-prediction backbone with time embeddings, the training objective, and the reverse sampling loop.The project was written as a way to work through the mechanics of generative diffusion in detail, and doubles as a readable reference implementation.

Twitter Job Title Prediction

2019
Associated with Adak AI

Python NLP job title prediction

Python NLP, job title prediction, word embeddings, TF-IDF, web scraping, semantic clustering, classification, deep learning, inference The more social media take its place in our lives; the more critical their analysis becomes and the more researchers’ attention is drawn to it. Studies contain various topics such as sentiment analysis, trend prediction, bot detection, Etc. Here, for the first time, we propose a novel method to predict the job title of social media users. Twitter, a popular social media, is our target social media. We introduce a dataset consisting of 1314 samples, including users’ tweets and bios. The user’s job title is found using Wikipedia crawling. The challenge of multiple job titles per user is handled using a semantic word embedding and clustering method. Then, a job prediction method is introduced based on a deep neural network and TF-IDF word embedding. We also use hashtags and emojis in the tweets for job prediction. Results show that the job title of users in Twitter could be well predicted with 54% accuracy in nine categories.

YouTube Downloader

2019
Associated with Communication and Computation Lab, University of Tehran

Python PyQt YouTube downloading