💻Mohammadhossein Akbari Moafi
👋 Hi, there! I’m Mohammadhossein (Rastin) Akbari Moafi, an AI Researcher & Engineer.
Check out my resumé and portfolio below.
Publications
Projects
Translation RAG Service
A lightweight Retrieval-Augmented Generation (RAG) backend for translation prompts. The service exposes a FastAPI REST API that stores translation pairs in SQLite — kept idempotent through database-level deduplication — retrieves the most relevant examples with TF-IDF cosine similarity, and assembles them into retrieval-augmented prompts for LLM-based translation. An additional heuristic component detects stammering in translated sentences, and the whole system ships Dockerized for easy deployment.
KG-Research-Agent
Built a multi-agent, evidence-grounded research system that ingests scientific PDFs, retrieves context with RAG (ChromaDB + Gemini), extracts structured claims and evidence, and stores them in a Neo4j knowledge graph. Features a Planner → Retriever → Evidence → Answer pipeline with session memory, citation-grounded responses, and end-to-end PDF ingestion, querying, and multi-turn reasoning.
Support Memory Weave
Built an LLM-powered pipeline for transforming unstructured customer support messages into structured tickets using FastAPI, LangGraph, and PostgreSQL. Features automated conversation ingestion, multi-step ticket structuring, semantic vector retrieval with SentenceTransformers, and evidence-based reply suggestions backed by realistic conversational datasets.
Provenance-Aware Metadata (Phase 3)
Developed a provenance-aware metadata workflow for digital assets using JSON-LD (Dublin Core, PROV-O, Schema.org) with SHACL validation, C2PA signing, and a FastAPI service exposing IIIF Presentation 3.0 manifests and verification endpoints. Implemented Wikimedia Commons integration, extended EDM/PREMIS validation policies, CI/CD pipelines, and container publishing via GHCR.
Reviews Dashboard
Built a modern dashboard for property managers to analyze and manage guest reviews across properties, integrating Hostaway and Google Reviews APIs with trend analysis, approval workflows, modular UI components, and extensible LLM-based review analysis for sentiment and insights.
EEG Motor Imagery – Deep ERP Classification
Processed ERP segments from the PhysioNet EEG dataset (109 subjects) for motor imagery classification. Evaluated RF, LSTM, CNN, EEGNet via cross-validation and full-dataset training. Reached 74.34% accuracy with EEGNet, outperforming traditional and sequence models.
3D Explainable AI for Renal Structure Segmentation in CT Scans
An explainable AI project for renal structure segmentation in CT imaging. SegResNet was fine-tuned on multi-phase renal CT scans, reaching a Dice score of 0.78 for organ segmentation. To make the model’s decisions transparent, 3D Grad-CAM and Guided Backpropagation were integrated to produce volumetric, class-specific visualizations that show which anatomical regions drive each prediction.
Knowledge-Guided Visual Analytics for Scientific Image Annotation
A deep learning pipeline for automatic annotation and retrieval of scientific figures. The system builds on CLIP and BiomedCLIP — including fine-tuned variants — to embed biomedical images and text into a shared semantic space, and integrates Wikidata-based knowledge graphs to enrich that space with structured domain knowledge. The result is stronger semantic alignment between figures and their descriptions, along with interpretable visualisations that make the annotation and retrieval behaviour of the models easier to inspect.
Text-Guided Medical Image Denoising with Vision-Language Fusion
A medical image denoising system that fuses vision and language: a UNet-based denoiser is conditioned on CLIP text embeddings derived from the VQA-RAD dataset, letting textual context guide the reconstruction of noisy scans. Compared to a vision-only baseline, the vision-language fusion improved reconstruction quality by 12.5% in PSNR and 4.6% in SSIM.
Football Match Dynamics - Centralities, Motifs, and Embeddings
Analyzed passing networks with graph-based metrics and Node2Vec embeddings. Identified tactics and key players using Louvain clustering and temporal trends.
Stock Price Prediction
Designed an LSTM-based model for 14-day forecasting with real-time data pipelines. Improved financial modeling for short-term predictions.
Sport Video Analysis for Billiard Matches
Built a real-time tracking system in C++ with OpenCV for cue ball and table detection. Visualized ball trajectories across angles using dynamic 2D minimaps.
PointNet-SHOT - Enhanced 3D Descriptor Matching
Enhanced PointNet with SHOT descriptors and custom triplet loss. Achieved 94.12% accuracy for robust 3D object recognition on noisy datasets.
Lymphoma – Data Augmentation for Deep Medical Imaging
Boosted DenseNet121 classification accuracy to 98.67% via PCA, DCT, and noise-based augmentation (baseline - 93.6%).
Optimized Point Cloud Alignment
Implemented ICP with KD-Tree and SVD for faster, precise alignment in industrial 3D reconstruction tasks.
3D Scene Reconstruction Using Structure from Motion
Built a 3D reconstruction pipeline using ORB matching, RANSAC, and geometric validation from multi-view images.
Semi-Global Stereo Matching with Monocular Disparity Guess
A stereo depth estimation project that combines classical and learning-based approaches: a monocular disparity prediction is used as the initial guess, which Semi-Global Matching (SGM) then refines into the final disparity map. Seeding SGM with the monocular estimate substantially improved accuracy over the baseline, reducing mean squared error by 75%.
TiagoBot Navigation and Pick-Place System
Programmed ROS1-based routines for autonomous movement and pick-place using AprilTag recognition.
Bone Age Prediction from Hand Radiographs
Used MediaPipe and CLAHE for preprocessing. Achieved 10.47 MAE using deep regression (ResNet, Inception-v4).
Used Car Price Prediction
An end-to-end machine learning pipeline for estimating the market price of used cars. Listings were collected via web scraping and cleaned into a structured dataset, with categorical features encoded for model consumption. Several regression models were trained and compared, with a Random Forest regressor delivering the most accurate price estimates and forming the core of the final predictive system.
Tehran Air Quality Index Forecasting
A forecasting study of Tehran’s Air Quality Index built on a decade of historical pollutant measurements. Decision Tree, Support Vector Machine, and Random Forest models were trained and evaluated on the dataset to predict AQI levels, comparing how well each captures the temporal patterns and pollutant interactions that drive air quality in the city.