Skip to content
View Nikelroid's full-sized avatar
🎯
Focusing
🎯
Focusing
223 followers · 204 following

Achievements

Achievement: YOLOAchievement: Pull Shark

Achievements

Achievement: YOLOAchievement: Pull Shark

Block or report Nikelroid

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Nikelroid/README.md

Hi 👋, I'm Nima Kelidari

MS in Computer Science (AI) @ USC Viterbi

Large Language Models · Multimodal Learning · Numerical Optimization

🎓 About Me

  • 🏥 Now (Summer 2026): Computational Optimization Intern at Memorial Sloan Kettering Cancer Center — GPU-accelerated convex solvers behind radiotherapy treatment planning (~2× faster warm starts, ~100× tighter optimality gap than the production solver)
  • 🔭 MS in Computer Science (AI) at USC Viterbi, finishing Dec 2026 (GPA 3.55)
  • 🎓 BS in Computer Science from Sharif University of Technology (Aryamehr University of Technology)
  • 🧪 Running three concurrent research projects with USC NLP, USC ICT (IHP Lab), and Prof. Yan Liu's lab
  • 📄 First preprint on arXiv — 2607.06854 — and an MoE-steering manuscript in preparation
  • 💼 Open to summer research and PhD positions for Fall 2027
  • 📍 Los Angeles, CA

🔬 Current Research

🧠 3D-DELTA — Learned Low-Rank Steering for MoE LLMs

With Profs. Robin Jia & Xuezhe Ma · USC NLP

A 50K-parameter learned head replacing SteerMoE's hand-engineered statistic for routing intervention on OLMoE-1B-7B. Beats SteerMoE (ICLR 2026) by +0.101 mean across six faithfulness benchmarks — 4× its gain — with +0.288 on FaithEval-Inconsistent, and the recipe transfers across three MoE families (7B–47B, 8–128 experts). Manuscript in preparation. · Project page

🎮 Adversarial Co-Evolution of RL and LLM Agents

With Prof. Yan Liu · USC · 📄 arXiv:2607.06854

Distributed PPO/TRPO pipeline with a rising-opponent curriculum over 100+ controlled runs, reaching a 99.12% win rate vs baseline agents on Gin Rummy. Master–Worker inference engine (14 GPU workers, ~32 queries/s, 62× faster model loading) bringing Llama-3 / Gemma / GPT-OSS into the training loop. Project page

🎙️ Linguistic-Agnostic Speech Emotion Recognition

With Prof. Mohammad Soleymani · USC ICT (IHP Lab)

MLOps probing framework in PyTorch + HuggingFace over six pretrained speech encoders × eight datasets via SLURM, with a FastAPI / JS dashboard for real-time layer-wise paralinguistic analysis. Now extended with a sparse-autoencoder dissection: SAEs explaining 93% of the signal over ~940K frame activations isolate 61 confound-controlled emotion features (60/61 acoustic — the models listen to tone, not words). Project page

Fun: 🎸 Dastan — Gesture Guitar Studio

Play real guitar with bare hands and a webcam: gesture-controlled chords, Persian & flamenco rhythms, a theremin-style melody hand, and in-browser recording — all in one dependency-free HTML file. ▶ Try it live

More projects — from an 80M-interaction recommender on GKE to SVD image compression — in the full index.

🛠️ Tech Stack

Languages

Python Java C++ SQL R

ML / Deep Learning

PyTorch JAX TensorFlow HuggingFace vLLM scikit-learn Keras OpenCV NumPy Pandas

GPU, HPC & Optimization

CUDA CuPy SLURM ADMM Interior--Point

MLOps & Infrastructure

Docker Kubernetes Jenkins MLflow Weights & Biases FastAPI

Databases & Cloud

MySQL PostgreSQL MongoDB Redis Google Cloud AWS

Robotics

ROS Kinematics Motion%20Planning

📊 GitHub Stats

Nima's GitHub Stats Top Languages

📫 Get in Touch

🌐 Website: kelidari.com  ·  📧 Email: kelidari@usc.edu  ·  💼 LinkedIn: nima-kelidari  ·  🎓 Scholar: Nima Kelidari

Thanks for visiting! ⭐

Pinned Loading

  1. adversarial-coevolution adversarial-coevolution Public

    Adversarial Co-Evolution of RL and LLM Agents: A framework for training high-performance PPO agents against Large Language Models in Gin Rummy, utilizing curriculum learning and knowledge distillat…

    Python 2

  2. multimodal-sentiment-classification multimodal-sentiment-classification Public

    A multimodal deep learning framework that fuses visual features from EfficientNet and textual features from BERT to classify sentiment in image-text conversations using the MSCTD dataset.

    Jupyter Notebook 9 2

  3. anime-recommender-application anime-recommender-application Public

    An end-to-end MLOps project implementing a deep learning-based anime recommendation system with automated CI/CD deployment to Google Kubernetes Engine.

    Python

  4. image-compression-svd-fft image-compression-svd-fft Public

    A Python implementation of lossy image compression and reconstruction utilizing Singular Value Decomposition (SVD) and Fast Fourier Transform (FFT) techniques for optimal storage efficiency.

    Jupyter Notebook 3

  5. artist-explorer-app artist-explorer-app Public

    A comprehensive full-stack web and Android application for exploring artists and artworks using the Artsy API, featuring JWT authentication, MongoDB favorites, and an experimental AI assistant.

    HTML

  6. mysql-metadata-manager mysql-metadata-manager Public

    A Python-based GUI application using Tkinter and MySQL to reverse-engineer, manage, and export database schema metadata.

    Jupyter Notebook 3