Hybrid Anime Recommender System
Neural collaborative filtering over 80M+ user-anime interactions (300K users × 14K titles), with Comet ML and DVC tracking, Jenkins CI/CD, and deployment on GKE.
Catalogue · 2021 — present
Everything, or close to it: coursework that got out of hand, research prototypes, weekend builds, and a few things made purely for the fun of it. The six projects I'd show first are on the home page; the rest live here, grouped by area. Nearly all of it is public on GitHub.
End-to-end ML systems with production pipelines, deployment, and experiment tracking.
Neural collaborative filtering over 80M+ user-anime interactions (300K users × 14K titles), with Comet ML and DVC tracking, Jenkins CI/CD, and deployment on GKE.
LightGBM + Flask application predicting hotel booking cancellations, with a full MLOps pipeline from data ingestion to deployment.
Compact student networks trained via distillation from larger teachers, keeping accuracy while cutting compute.
Comparative study for image classification with ablations over dropout, kernel factorization, and architecture choices.
Hand-built network trained with SGD on Fashion-MNIST, evaluated across multiple optimization regimes.
Neural network forecasting Iran's group-stage results from historical match data.
MLP for Latin-letter recognition from hand-movement data, compared across SGD and Adam, validated on live webcam input.
Latent-factor and SVD-based methods predicting user preferences from sparse interaction data.
Reference implementations: music recommendation, heart-disease classification, regression models.
Sequence modeling from Persian-language tasks to image captioning.
CNN feature extractors feeding LSTM/GRU decoders for automatic caption generation.
Sequence models generating verse in the style of Ferdowsi, trained on a Persian poetry corpus.
Attribution of poems across ten Iranian poets with high accuracy on held-out verses.
Curated, preprocessed corpus with stopword lists, ready for generation and classification tasks.
Classical and deep approaches: detection, segmentation, stitching, reconstruction.
Scene classification across AlexNet configurations: transfer learning and feature-extraction strategies.
SIFT descriptors and BoVW pipeline, with epipolar-geometry analysis for multi-view scenes.
Corner detection and homography-based stitching with RANSAC for robust panoramas.
Histogram-of-oriented-gradients features with an SVM classifier and non-maximum suppression.
Panoramas from video sequences plus background subtraction, via SIFT-based homography estimation.
K-Means, Mean-Shift, SLIC superpixels, and GrabCut, implemented and compared on real data.
Face morphing, Poisson blending, and Laplacian stacks for seamless composites.
Full-color images recovered from historical black-and-white plates by pyramid-based RGB channel alignment.
Histogram matching between images to enhance detail and tonal quality.
Multi-scale template matching in color, with edge detection and match highlighting.
Rectifying book covers from photographs: corner selection, homography computation, warping.
High-pass/low-pass frequency-domain blends: one image up close, another from across the room.
Line and corner detection by parameter-space voting after edge detection.
Patch-based synthesis with Dijkstra-optimal seams and gradient handling.
Filling missing regions with patch-based synthesis followed by smoothing.
Web and mobile applications, backend to frontend to database.
One product, three builds: Flask + vanilla JS, Angular + Node.js + MongoDB Atlas, and Android in Kotlin with Jetpack Compose. Artist search, auth, and favorites on the Artsy API. (Code private — ask for access.)
SwiftUI client for the Real-Time AI Vision Assistant: camera streaming, animations, activity suggestions.
Tkinter GUI that reverse-engineers and manages database schema metadata, with import/export and row-level editing.
Low-level systems, distributed networking, and classical algorithms.
Multi-threaded torrent-style file sharing over TCP sockets with a central tracker, peer discovery, and detailed logs.
Custom eight-switch topology with a Floodlight controller doing K-shortest-path routing.
Custom Java thread pool: resource-aware scheduling, priority queues, dynamic worker scaling.
Stack machine in MIPS assembly with expression evaluation and a custom calling convention.
Lossless and lossy text compression with frequency analysis and tree serialization.
Linear-grammar parsing and DFA minimization with automaton visualization.
Computational geometry, graph theory, and dynamic programming problem sets.
Curated solutions spanning competitive programming and academic problem sets.
Games and toys exploring AI, multiplayer architecture, and engine fundamentals.
Strum real guitar chords with hand gestures through your webcam: Karplus-Strong synthesis, Persian 6/8 and flamenco rhythms, a theremin-style melody hand, and recording — one dependency-free HTML file. code
Client-server battleship with accounts, hosted matches, spectator mode, and a leaderboard.
JavaFX clone of the arcade classic: power-ups, save/load, scoreboard, many brick types.
Pac-Man-style game in C++ with BFS ghost AI, level loading, and a custom render loop.
A multi-component social platform: backend, desktop clients, CLI, and chatbot extensions.
Java backend with MongoDB, authentication, timelines, and a REST API for clients and bots.
JavaFX desktop client with offline support and message caching.
Alternative JavaFX client focused on UI animation and component composition.
Terminal client with JSON persistence and offline composing.
Tic-tac-toe over messages, with dynamic board rendering.
Expression-parser bot with an extensible operator set.
Interactive polls with real-time tallies.
Statistical modeling and exploratory analysis in R.
Excess-mortality estimates across 31 provinces: regression on pre-COVID data versus observed deaths.
Decision trees, random forests, neural networks, and KNN compared on a balanced diabetes dataset.
Older odds and ends live in the full repository list on GitHub.