Projects

Magics.AI

Machine Learning(ML) System, Fine-Tuning, Large Language Model(LLM), K8S, Docker, Frontend, Backend

Magics.AI is an open-source platform designed for the academic community, providing tools for fine-tuning and inference of large language models (LLMs) with reduced costs and latency. It supports distributed resource integration across institutions, features a Python SDK and user-friendly interface, and lowers the technical barrier for model fine-tuning. Additionally, Magics.AI supports Embodied AI, enabling LLMs to control robotic systems and interact with virtual environments, bridging language understanding and physical actions for interdisciplinary research.

Merry Query

Large Language Model(LLM), Retrieval-Augmented Generation(RAG), Agent, Frontend, Backend

MerryQuery is an AI-powered educational assistant that utilizes retrieval-augmented generation (RAG) to provide students with tailored responses based on course materials.This tool is designed to support both teachers and students. Teachers can input course materials, and define data controls to prevent undesirable content generation. Students receive responses tailored to their previous interactions and course materials with references. You can visit our website for more information: https://exploremq.benyamintabarsi.com/

Gentopia.AI

Large Language Model(LLM), Retrieval-Augmented Generation(RAG), Agent, Tool Call

Gentopia is a lightweight and extensible framework for LLM-driven Agents and ALM research. It provides essential components to build, test and evaluate agents. At its core, Gentopia aims to assemble an agent with a single config, thus minimizing your effort in building, tuning, and sharing agents.

GentBench

Large Language Model(LLM), LLM Evaluation, Benchmark

Benchmark and Evaluation for Gentopia. A good ALM Benchmark should aim to solve problems and tasks hard/unsolvable by LLMs (else there is no meaning to pay tool tax). Tasks in Gentbench will be half-public and half-private. We open-source a demonstrative public benchmark to encourage agent tuning, but will use a private bench (of similar distribution) for fair eval.

SageCube

Large Language Model(LLM), Retrieval-Augmented Generation(RAG), Agent, Tool Call, Electron, Agent Platform

SageCube is an AI assistant designed for desktop use and is currently being tested on Steam. Powered by an LLM, SageCube offers a visually appealing user interface. Users can interact with interactive Live2D and 3D virtual avatars using voice or text. SageCube supports various voice models for text-to-speech functionality, enabling users to interact primarily through voice. Additionally, through Steam’s Workshop, users can upload and install various tools and acquire agents, enhancing the functionality and customization of SageCube.

SUSTech Online Judge (OJ) System: JCoder

Security, Remote procedure call(RPC), K8S, Docker, Sandbox, Database, Frontend, Backend

JCoder a scalable online judge system to evaluate code correctness across multiple programming languages including C/C++, Java, JUnit, SQL, Python, MIPS and Verilog. The system passed third-party penetration testing and is now officially used by the Computer Science Department at Southern University of Science and Technology (SUSTech) , serving over 3,500 students in 13 courses.

Multi Person Conference, Remote Control System

Network, Electron, Frontend, Backend, Multi-Platform, P2P

Design and implement a cross-platform remote multi-person conference, chat, remote control system. I took charge of all parts, used QUIC protocol to encapsulate the packets, Electron for multi-platform frontend, and applied P2P for decentralized permission verification This project won the highest grade in the class!

Domain Specific Language for Online Judger

Domain Specific Language(DSL), Security, Remote procedure call(RPC), K8S, Docker, Sandbox, Rust, Compiler

The Domain Specific Language for Online Judger is a DSL implemented using macros in Rust. It allows users without any Rust or programming knowledge to define evaluation tasks using simple statements. This DSL supports a variety of functions, including defining file systems, setting up runtime environments (such as nsjail, Docker, VMs), and specifying tasks. The compiler translates this DSL into Rust code for execution, ensuring high operational efficiency.

Heterogeneous Bert

Bert, Neural Architecture Search(NAS), Knowledge Distillation, Evolutionary Algorithm(EA), Super-Network

We have implemented a neural architecture search and a super-network training framework for heterogeneous BERT models. Given the search space and a teacher model, the super-network is automatically trained and the network structures are evaluated using balanced Pareto sampling. Compared to traditional neural architecture search frameworks, our approach achieves higher accuracy, faster convergence for sub-models, and superior performance under the same structural configurations.

EvoXbench: All-In-One Neural Architecture Search Framework

Neural Architecture Search(NAS), Benchmark, Evolutionary Algorithm(EA)

EvoXbench is an open-source library that serves as a comprehensive framework for the development of Neural Architecture Search (NAS) algorithms. It integrates all essential technologies and provides user-friendly Python and MATLAB interfaces, enabling the easy testing and development of algorithms. Key tasks included the collection, extraction, and curation of extensive NASBench datasets using Django’s ORM framework, and training an MLP surrogate model using PyTorch while overseeing the experimental processes.

SUSTech Library Webpage

Data Mining, Database, Frontend, Backend

The SUSTech Library Webpage utilizes data mining and analysis techniques to compile and analyze borrowing and access data from faculty and students across the entire school. This results in the production of reports that help the school optimize the borrowing experience. Additionally, personalized user profiles are provided for each user.

GroceryHelper

Firebase, Computer Vision(CV), Frontend, Backend

GroceryHelper is a mobile app that enables users to identify the quantity and expiration dates of groceries by simply uploading photos. The app intelligently recommends recipes based on the shelf life of the groceries and sends email reminders before the items expire. This helps users manage their pantry efficiently and reduce food waste by suggesting timely culinary uses for the ingredients they have on hand.

Reversed Reversi(Alpha Zero)

Alpha Zero, Reinforcement Learning(RL), Evolutionary Algorithm(EA), Monte Carlo Tree Search(MCTS)

Designed an AI for reversed reversi, and used alpha zero to build the AI model. Also applied evolutionary algorithm to optimize the weight in the reinforcement learning training network. The model won the 1st place in SUSTech in the grade with 402 wins, 1 loss and 1 draw, leading significantly in winning rate.