Nima Kelidari

Memorial Sloan Kettering Cancer Center, Department of Medical Physics

• IMRT radiotherapy planning reduces to a convex quadratic program with up to ~216K beamlet variables per clinical case. I work on making that solve fast: a path-following ADMM that warm-starts an interior-point method, GPU-accelerated with CuPy and mixed precision on V100 and P100 nodes.
• Implemented and benchmarked first-order methods from the recent literature — PDHG/PDQP (JAX/MPAX), ADMM with over-relaxation and adaptive penalties (residual-balancing, OSQP-style, spectral/ARADMM), Fast-ADMM with restart, and safeguarded Anderson acceleration — on real lung and prostate cases.
• Replaced the solver's hand-tuned settings with Bayesian optimization over 1,015 Optuna trials: 78 → 21 iterations to a usable warm start on a held-out clinical case — 3.7× fewer, 3.2–3.8× less wall-clock — then derived and validated a bound that explains the gain rather than just reporting it.
• Rebuilt the penalty recovery rule so runtime barely depends on a mis-set starting penalty — 6 → 21 iterations across four orders of magnitude, against 8 → 78 for the prior rule and 20 → 337 for the textbook one — on a platform behind ~1,100 solver runs: a matrix-free warm-started conjugate-gradient GPU engine, an NFS-safe SLURM search harness, and an early exit cutting trial cost 1.8–4.4×.

University of Southern California · Los Angeles

• Inference-time steering for mixture-of-experts LLMs — a 197K-parameter learned low-rank head that intervenes on MoE routing with no changes to model weights. Trained on 9,027 contrastive pairs from SQuAD, HotpotQA, and FEVER, it improves mean faithfulness by +0.101 across six benchmarks — 4× the gain of the published SteerMoE baseline (ICLR 2026) — and the plugin runs unmodified across four MoE families (1.3B–47B, 8–128 experts), with results reported on OLMoE. Sole-authored manuscript in preparation.
• Adversarial co-evolution of RL and LLM agents — a distributed PPO pipeline with a three-phase curriculum reaching a 99.12% win rate against baselines, plus a master–worker inference engine that distills strategy from Llama-3, Gemma, and GPT-OSS into compact RL policies. First-author paper accepted to AIIDE 2026 (AAAI) — arXiv:2607.06854.
• Language-agnostic speech emotion recognition, with Prof. Mohammad Soleymani (USC ICT, IHP Lab) — a PyTorch / Hugging Face probing framework evaluating six pretrained speech encoders across eight datasets on SLURM, with a FastAPI dashboard for layer-wise paralinguistic analysis. A sparse-autoencoder dissection explaining 93% of the signal over ~940K frame activations isolates 61 confound-controlled emotion features, 60 of them acoustic — the models listen to tone, not words.

Sharif University of Technology

• Ten months in Dr. Tefagh's group on a four-person team, surveying and prototyping generative computer-vision pipelines built on vision transformers and 3D CNNs.
• Ran the experimental side across Git, Kaggle, Google Colab, and university GPU servers, keeping the team's long-running training jobs organized and reproducible.

Sharif University of Technology

• Four semesters, six courses, 300+ students — lectures, assignment design, quizzes, and grading for basic and advanced programming, computer fundamentals, and machine learning, working in TA teams alongside head TAs and course professors.
• Built and delivered TensorFlow workshops and diagnostic sessions for sixty students in the ML course, holding 65% attendance across the series.

University of Southern California, Viterbi School of Engineering · Los Angeles

• GPA 3.55 / 4.0. Coursework in applied NLP, multimodal probabilistic learning, machine learning, deep learning, and web technologies — with NLP and robotics next, to match where the field is going.
• Off the clock: volleyball and hiking.

Sharif University of Technology · Tehran

• GPA 3.37 / 4.0 overall, with fifteen-plus CS and AI courses at a 3.5 / 4 average — computer vision, deep learning, machine learning, algorithm design, numerical analysis, data structures, image processing, databases — on top of a heavy mathematical core.
• Ten months of computer-vision research in Dr. Tefagh's group; teaching assistance and the volleyball court filled the rest.