Work
Achievements & Honors
Google DeepMind APAC Research Symposium 2026
Selected as a student research delegate.
2026
EY Techathon 6.0
Secured a Top-3 finish among 180,000+ participants nationwide.
2026
Amazon ML Challenge 2025
Ranked 66th out of 82,789 participants in a national ML competition.
2025
STC-UP Hackathon
Ranked Top 11 across India; only undergraduate team competing against industry startups.
2025
Research & Professional Experience
Research Intern
Causal Evidence & Decisions Studio, Prof. Harsh Parikh – Yale University
May 2026 – Jul 2026
New Haven, USA
New Haven, USA
- Formulating a theoretical framework to construct Confidence Sets of Representations.
- Implemented and evaluated multiple causal representation learning algorithms, applying structural relaxations to bound the unidentifiable representation space.
- Overcoming the limitations of purely predictive models by recovering true causal mechanisms, enabling robust counterfactual reasoning and scientific discovery.
Undergraduate Research Assistant
P-Square Lab, Prof. Parikshit Pareek – IIT Roorkee
Sep 2025 – Present
Roorkee, India
Roorkee, India
- Demonstrated orthogonal task structures induce disentanglement in MLPs, enabling surgical unlearning. [Read Paper]
- Designed amortized inference for zero-shot Gaussian Process kernel discovery from limited data using PFNs.
- Engineered a sparsity-aware latent diffusion model for ultra-sparse jet image synthesis in high-energy physics (CERN). [Read Paper]
Research Engineer
Elimentary (YC F26)
Nov 2025 – Jan 2026
Bangalore, India
Bangalore, India
- Built secure, quantized local LLM APIs, achieving granular layer-wise observability.
- Optimized GPU throughput by implementing sequence scheduling for concurrent LLM inference requests.
- Designed a no-code LLM evaluation platform integrated with Langfuse and Grafana for comprehensive monitoring.
Publications
-
Structural Disentanglement in Bilinear MLPs via Architectural Inductive Bias
ICLR Under Review Under review at ICLR 2027 2026Architectural inductive bias supports structural disentanglement which, in turn, helps unlearning and generalization. -
-
Localised Routing, Distributed Execution: How LLMs Process Semantic Domains
ICLR Under Review Under review at ICLR 2027 2026Investigating how LLMs internally represent and process diverse domain knowledge. Found that attention routes domain identity while MLPs store domain-specific logic. -
Image-Alchemy: Personalized Text-to-Image Generation
ICLR Workshop ICLR DeLTa Workshop 2025A personalization method for diffusion models advancing subject fidelity in text-to-image generation. -
Sparsity-Aware Representation Learning for Jet Image Generation via Guided Latent Diffusion
Preprint Preprint 2025Modeled a physics-aware diffusion model for jet image synthesis at CERN’s Large Hadron Collider.