cv
You can download the PDF from the pink/blue link on the right.
Basics
| Name | Kaustubh Sharma |
| Label | Undergraduate Student |
| kaustubh_s@ee.iitr.ac.in | |
| Url | https://github.com/kaustubh202/ |
| Summary | Electrical Engineering B.Tech student at IIT Roorkee with interests in mechanistic interpretability, diffusion models, and ML for science and engineering. |
Work
-
2025.05 - Present Research Assistant
P-square Lab, IIT Roorkee
Working on attention mechanism for Prior Data-Fitted Networks (PFNs).
- GP inference time reduction for physical equation solving
- Scaling of PFNs
-
2025.03 - 2025.03 Educator
Edufabrica Pvt. Ltd.
Delivered a 2-day workshop lecture to 200+ students across India on Generative AI.
-
Incoming Quantitative Research Intern
Goldman Sachs
Selected for the quantitative strategist internship involving financial modeling and statistical analysis.
Education
Awards
- 2025.01.01
Micron AI Hackathon 2025 — Second Runner Up
Micron
Achieved second runner up position at Micron AI Hackathon 2025.
- 2023.06.01
- 2023.05.01
- 2021.01.01
- 2022.01.01
Publications
-
2025 Image-Alchemy: Advancing Subject Fidelity in Personalized Text-to-Image Generation
DeLTa Workshop, ICLR
Designed pipeline for personalizing Stable Diffusion XL with high-fidelity subject integration, reduced overfitting and forgetting using LoRA and segmentation-guided Img2Img.
-
2025 Explainable AI-Generated Image Forensics: A Low-Resolution Perspective with Novel Artifact Taxonomy
APAI Workshop, ICCV
Developed robust interpretable pipeline for AI-generated image detection at 32×32 resolution with artifact taxonomy and explainability framework.
Skills
| Machine Learning | |
| Mechanistic Interpretability | |
| Diffusion Models | |
| ML for Science and Engineering |
Languages
| Hindi | |
| Native speaker |
| English | |
| Advanced |
| French | |
| Beginner |
Interests
| Music | |
| Piano |
| Swimming | |
| Competitive Swimming |
| Data Science | |
| Data Science Group — IIT Roorkee |
Projects
- 2025.04 - Present
Domain Circuit Discovery in LLMs - Mechanistic Interpretability
Investigating domain-specific knowledge emergence in LLaMA 3-3B; mapping domain-specific 'rooms' across transformer architecture.
- Forward pass profiling
- Probe separability
- Zero-out tests
- Hydra effect
- Fine-tuning shifts
-
Sparsity-Aware Representation Learning for Jet Image Generation via Guided Latent Diffusion
Developed sparsity-aware latent diffusion for jet images in high-energy physics with novel VAE and mean-pulling mechanism.
- Custom VAE with sparsity reconstruction loss
- Latent diffusion mean-pulling mechanism