Skip to content
View Nikelroid's full-sized avatar
🎯
Focusing
🎯
Focusing

Organizations

@paralinguistics

Block or report Nikelroid

Report abuse

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

Report abuse
Nikelroid/README.md

Nima Kelidari

MS Computer Science (AI) @ USC Viterbi · graduating Dec 2026

Large Language Models · Multimodal Learning · Numerical Optimization

GitHub Stats Top Languages


About

  • 🏥 Optimization intern at Memorial Sloan Kettering — GPU convex solvers behind radiotherapy planning (78 → 21 iterations, 3.7× faster)
  • 🔬 Graduate researcher at USC — LLM steering, speech models, RL agents
  • 📄 First-author paper at AIIDE 2026 (AAAI) · arXiv:2607.06854 · co-author preprint arXiv:2608.04669
  • 🎓 BS Computer Science, Sharif University of Technology
  • 💼 Open to full-time roles from Jan 2027, and PhD positions for Fall 2027

Research

  • 3D-DELTA — a 197K-parameter head steering MoE routing at inference with zero weight changes, 4× the SteerMoE (ICLR 2026) gain; transfers across four MoE families, 1.3B–30.5B
  • Adversarial Co-Evolution — masked PPO/TRPO against a gold-standard expert and distributed LLM opponents, on PettingZoo/RLCard · AIIDE 2026
  • Language-Agnostic SER — 6 speech encoders × 8 datasets; sparse autoencoders isolate 61 emotion features

Try it in your browser

No install, no signup. Everything below is live.

🎭 Tri-Modal Sentiment Playground Type, show your face, or record your voice — three models read you independently and a confidence-weighted fusion arbitrates. Local mode keeps your data on your machine
🎙️ Voice Calibration Lab Teach it your prosody in 5 minutes — 88.3% → 96.7% leave-one-out
😐 Face Calibration Lab Seven acted expressions, auto-captured — 81% → 90.5% leave-one-out
🧪 ML from Scratch — Interactive Lab 10 algorithms in NumPy, each training live in the browser: backprop, AdaBoost, K-means++, HMM Viterbi, attention, Q-learning
♟️ Co-Evolution Explorer All 76 runs / 412 GPU-hours behind the AIIDE paper — including the 99.12% that turned out to be 34%
🗣️ Speech Emotion Explorer Layer-by-layer probing across 6 encoders, 8 corpora, 5 languages
🎸 Dastan — Gesture Guitar Play chords with your hands. One HTML file, 25 ms audio lookahead
⚽ World Cup 2026 Schedule Built because I wanted it to exist

Stack

Python Java C++ SQL · PyTorch JAX TensorFlow Hugging Face vLLM · CUDA CuPy SLURM · Docker Kubernetes Jenkins GitHub Actions MLflow FastAPI · GCP (Cloud Run) AWS Azure · ROS

More in the full project index.

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 1

  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 10 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. 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

  5. dastan-gesture-guitar dastan-gesture-guitar Public

    🎸 Play guitar with your bare hands — webcam gesture chords, Persian & flamenco rhythms, a theremin-style melody hand, and in-browser recording. One dependency-free HTML file.

    HTML

  6. linguistic-agnostic-ser linguistic-agnostic-ser Public

    A probing framework for Speech Emotion Recognition (SER) transformers to evaluate how and where paralinguistic, acoustic knowledge is encoded across their hidden layers.

    Jupyter Notebook