Theoretical Physicist · Scientific ML Researcher · AI Systems Builder
I recently completed my PhD in theoretical physics at SISSA, where my research focused on three-dimensional mirror symmetry and dualities in supersymmetric quantum field theories. My doctoral work developed systematic methods for constructing dual pairs in a class of theories that had remained difficult to classify for decades, with findings published in Physical Review D and the Journal of High Energy Physics.
📄 Find my publications here
My interests now lie at the intersection of scientific machine learning, AI for Science, and AI systems engineering. I work on learning scientific dynamics, geometric deep learning, quantum error correction, and machine learning methods for scientific discovery.
I am interested not just in whether a model makes accurate predictions, but in whether it learns the underlying structure of a problem, generalises beyond its training distribution, and produces reliable results.
Alongside research, I build end-to-end AI systems: from model development and evaluation to APIs, containerised services, agentic workflows, and deployment. I enjoy working across the full stack, particularly when the application has a genuinely interesting scientific or technical problem behind it.
Languages
Data Handling
ML & Deep Learning
LLMs & Agentic AI
MLOps, Deployment & Tooling
Projects exploring how machine learning models represent, approximate, and generalise scientific systems.
Spectral Response Fidelity · A JAX/Equinox study investigating whether Fourier Neural Operators learn physically meaningful response functions, rather than merely accurate forward predictions. Explores Jacobian-supervised training, spectral response fidelity, differentiable quantum dynamics, and out-of-distribution generalisation.
Neural Operators for PDEs · An empirical study of how architectural inductive biases influence the data efficiency of neural operators.
- FNO reaches 10% test relative L2 error with approximately 4.1× less training data than a matched-budget U-Net.
- 116.3× faster inference than the finite-difference solver in the documented CPU benchmark.
- Includes multi-seed evaluation, bootstrap confidence intervals, resolution-transfer analysis, and spectral error decomposition.
Machine Learning Assisted Quantum Error Correction · A streaming GNN decoder for the rotated surface code under circuit-level SI1000 noise, exploring learned decoding strategies and correlated error information. Also extending the framework to color codes and investigating generalisation across code families. In progress
Building machine learning models and computational systems that connect scientific problems with practical, usable tools.
MolPredict · A molecular property prediction system using GATv2 on the QM9 dataset to predict HOMO-LUMO gaps.
- Test MAE: 0.1712 eV.
- FastAPI inference service with single and batch prediction.
- Dockerised CUDA-enabled deployment and an independent MCP server.
- MC-dropout-based uncertainty estimation. Live API documentation
Multi-Agent Symbolic Reasoning · An agentic system for mathematical problem-solving using a solver → verifier → critic architecture. Combines LangGraph orchestration with exact symbolic verification through SymPy, exposed as an independent MCP service. 🤗 demo
Building complete AI applications, with an emphasis on modular architecture, evaluation, observability, and deployment.
Agentic AutoML · An end-to-end agentic system for autonomous supervised machine learning.
- LangGraph orchestration and modular MCP tool server.
- FastAPI backend and frontend integration.
- Five-container Docker architecture.
- MLflow experiment tracking and LangSmith observability.
NL-to-SQL Agent · An agentic natural-language-to-SQL pipeline using hybrid retrieval and execution-guided self-correction.
- 54.5% execution accuracy on BIRD Mini-Dev (36/66).
- Approximately 55% fewer schema prompt tokens compared with full-schema prompting.
- Retrieval ablations, failure-mode analysis, and execution-guided correction. 🤗 demo
Exploring how models can be adapted to specialised domains, evaluated under challenging conditions, and made more reliable.
Domain-Adaptive LLM Trilogy · A three-stage exploration of language models for theoretical physics:
- RAG - retrieval over specialised physics knowledge.
- GraphRAG - structured knowledge representations and graph-based retrieval.
- QLoRA - domain adaptation of Qwen2.5-3B using a physics-specific instruction dataset.
Low-Resource NMT · Fine-tuning MarianMT for English-to-Malayalam translation using BPCC and NPTEL Shiksha data, with an emphasis on technical vocabulary and domain adaptation.
Multilingual LLM Safety & Red-Teaming · Investigating multilingual prompt injection and refusal robustness, with a particular interest in whether refusal behaviour relies on shared internal computations across languages and whether interventions can generalise to languages unseen during training.
Reinforcement Learning for multi-player strategy games · Graph neural networks and reinforcement learning, exploring GATv2 policies, action masking, differentiable graph construction, and curriculum learning.
Time Series Forecasting · LSTM, Transformer, AutoTFT and XGBoost across BTC/USDT and Indian electricity-demand forecasting, including climate stress tests.

