cv
Below is the condensed version of my CV. For the full CV please click on the Download PDF button.
Basics
| Name | Rishi Singhal |
| Label | Graduate Research Assistant |
| rsingha4@ncsu.edu | |
| Url | https://rishi2019194.github.io |
| Summary | PhD researcher in AI safety, interpretability, and alignment, specializing in analyzing memorization, privacy leakage, and bias in large-scale generative models (LLMs, multimodal transformers). I dissect and interpret model components at the neuron level using techniques such as quantization, pruning, and knowledge distillation to explain failure modes and biases, enabling design of scalable mitigation strategies. My work connects interpretability with efficiency to build robust and trustworthy AI systems in high-throughput, safety-critical environments. Available for Summer 2026 internships (May 2026 -- Aug 2026). |
Work
-
2024.05 - 2024.08 Machine Learning Intern
Fermi National Accelerator Laboratory (Fermilab)
Worked on deploying and optimizing machine learning systems for neutrino experiments.
- Deployed Graph Neural Networks (NuGraph2/3) on Fermilab’s EAF using Nvidia Triton & Docker.
- Enabled real-time background filtering and semantic labeling for MicroBooNE.
- Integrated Python/C++ client with LarSoft for direct streaming, reducing memory overhead by 20%.
- Extended NuSonic Triton framework for scalability and maintainability.
- Contributed production-level code adopted in Fermilab’s official reconstruction pipeline.
-
2024.01 - Present Graduate Research Assistant
Dr. Jung-Eun Kim Lab, North Carolina State University
Conducting research on memorization and generalization in deep neural networks, with a focus on transformer architectures.
- Discovered a novel role of LayerNorm in shaping memorization vs. generalization across Pre-LN and Post-LN models.
- Verified findings on both generative and classification tasks across NLP and CV.
- Showed that pruning only 0.1–0.2% of Post-LN parameters reduces memorization by ~70% without harming generalization.
- Demonstrated that early LayerNorms exert the strongest influence compared to later ones.
- Ongoing: Distinguishing memorization vs. generalization at the feature level and studying the impact of residual connections in large-scale LLMs (GPT, LLaMA).
-
2022.01 - 2023.04 Undergraduate Research Assistant
MIDAS Lab, IIIT Delhi
Conducted research on document coherence in NLP tasks.
- Investigated coherence as a core metric for evaluating text quality in summarization, translation, and QA.
- Applied Topological Data Analysis (TDA) on attention graphs of BERT, RoBERTa models.
- Developed lightweight MLP leveraging TDA features, outperforming transformer baselines by 5% on GCDC dataset.
Education
-
2025.01 - 2028.01 Raleigh, USA
-
2023.01 - 2025.01 Raleigh, USA
-
2019.01 - 2023.01 Delhi, India
BTech
Indraprastha Institute of Information Technology (IIIT) Delhi
Electronics and Communication Engineering
Publications
-
2025.09.18 Impact of Layer Normalization on Memorization and Generalization in Transformers
Accepted at NeurIPS 2025
Understand impact of Layernorm on memorization and generalization in Pre- and Post-LayerNorm transformers
-
2024.01.01 Beyond Words: A Topological Exploration of Coherence in Text Documents
ICLR 2024 (Tiny Papers Track)
Explores document coherence using topological methods.
Skills
| Programming & Tools | |
| Python | |
| C++ | |
| SQL | |
| MATLAB | |
| PyTorch | |
| TensorFlow | |
| Keras | |
| Scikit-Learn | |
| Numpy | |
| Pandas | |
| SpaCy | |
| NLTK | |
| Nvidia-Triton | |
| MCP | |
| Docker | |
| Flask | |
| Postman | |
| Git |
Projects
-
Exploring & Analyzing Internal Structure of Language Models to Mitigate Social Biases
Analyzed social biases in PLMs (BERT, RoBERTa), identifying their encoding in later layers and FFN, to inform pruning-based mitigation strategies.
- Bias analysis
- Model pruning
-
Few Informative Data Samples are Good Enough: Introducing Intelligent Data Pruning
Developed intelligent data pruning method for imbalanced datasets, outperforming SMOTE, Gaussian Copula, SDV, RRP.
- Data pruning
- Imbalanced learning