Experience
Physics and machine learning: robust analysis and development of generative models using statistical physics, and the use of these models to capture physics faithfully.
Research
May 2026 – Present Indian Institute of Science, Bangalore Project Associate (pre-doctoral researcher) Testing generative diffusion models on non-reciprocal turbulence, and following up my 2D XY study.
Generative diffusion models for non-reciprocal turbulence
- First application of generative diffusion models to non-reciprocal binary-fluid turbulence (2D Non-Reciprocal Cahn–
Hilliard– Navier– Stokes), benchmarking synthetic Lagrangian-tracer trajectories against direct numerical simulations. - Characterized Lagrangian multiscaling (velocity increments, their structure functions and flatnesses, multiscaling exponent ratios) for the first time in non-reciprocal hydrodynamics, alongside velocity and acceleration PDFs, curvature statistics, irreversibility and model transferability.
- Found that the accelerated DDIM sampler falters below 100 steps on some statistics, although it stays accurate down to 25 steps in 3D Navier–Stokes turbulence; an 800-step DDPM reconstructs the flow statistics accurately.
Follow-up to the 2D XY study (ongoing)
- Extending the study to finite-size scaling, and to diagnosing what the learned score network captures.
Jan 2026 – Apr 2026 Independent research Sole-author projects Two sole-author papers: diffusion on the torus for the 2D XY model across the BKT transition, and LLM-based verification of financial claims (3rd place, FMD Challenge, ICWSM 2026).
Manifold-aware score-based generative modeling of the 2D XY model
- Performed score-based diffusion directly on the torus (the 64 × 64 XY lattice, a 4096-dimensional torus); the learned score matches the theoretical Boltzmann score more precisely than a Euclidean-embedding diffusion model, which spends capacity learning the manifold constraint.
- To my knowledge, the first data-driven generative model (no explicit information of Hamiltonian, no feature engineering) to reproduce the heat capacity of the 2D XY model across the BKT transition to better accuracy than previously developed methods in literature; a MALA correction after ODE sampling removes the sampler-induced bias in the generated samples. Cv.
- Showed zero-shot generalization to other lattice size values from L = 8 to L = 64 without retraining, thereby capturing the BKT transition at unseen lattice sizes.
Courtroom-inspired verification of financial misinformation
- Addressed LLMs’ tendency to accept mathematically coherent but contextually false financial claims: an LLM (DeepSeek-chat) generates opposing prosecution and defense arguments for each claim, and a frozen ModernBERT encoder with parallel BiGRUs and context-attention pooling weighs the evidence.
- 3rd place on the FMD Challenge 2026 leaderboard: accuracy 0.955, F1 0.9549 on the blind test set.
Aug 2023 – Apr 2024 Indian Institute of Technology Roorkee Bachelor’s thesis researcher Transfer learning from nearest- to next-nearest-neighbor Ising models, a first step toward frustrated Josephson-junction arrays.
- Frustrated Josephson-junction arrays map onto long-range Ising models, which are costly to simulate with Monte Carlo; as a first step toward them, studied transfer from nearest-neighbor (NN) to next-nearest-neighbor (NNN) Ising models.
- Pre-trained deep networks on NN data and fine-tuned them on limited NNN data to predict phase transitions; compared fine-tuning configurations. Unlike earlier zero-shot NN→NNN transfer (Canabarro et al., PRB 2019; Corte et al., Comput. Mater. Sci. 2021), the network is adapted by fine-tuning.
- Earlier lab project (PHN300, 2023): classified Ising phases with feed-forward and convolutional networks trained on Monte Carlo configurations across lattice sizes.
May 2023 – Aug 2023 McGill University, Montreal MITACS Globalink Research Intern Monte Carlo particle-transport simulation (TOPAS, built on Geant4) of CR-39 detectors for neutron detection.
- Monte Carlo particle-transport simulation (TOPAS, built on Geant4) of CR-39 track-etch detectors for neutron detection: reproduced published recoil-proton threshold energies across nine aluminum degrader thicknesses in a radiator-degrader-detector setup, and compared secondary-particle dose deposition from an isotropic monoenergetic neutron source with and without a water phantom.
Industry
Jul 2024 – Dec 2025 OLA Krutrim, Bangalore Software Development Engineer I (AI Engineering & Research) Fine-tuned and aligned LLMs, and built production retrieval-augmented generation, model-serving and video-analytics systems.
A year and a half building and shipping AI systems in production: training and aligning large language models, retrieval-augmented generation, model serving and video analytics.
LLM fine-tuning and preference optimization
Fine-tuned LLaMA-3-8B-Instruct with LoRA on 1000+ code Q&A pairs for repository understanding. SimPO preference optimization reduced hallucination, with accuracy gains of 25% on training questions and 15% on paraphrased questions.
Retrieval-augmented generation platform
Built a production RAG platform with Flask, MilvusDB retrieval and Celery/Redis task queues.
Model-as-a-Service: Embedding and reranking models
Served embedding and reranking models on KServe/FastAPI at sub-200 ms latency.
Video analytics at scale
Built NVIDIA DeepStream CCTV analytics over RTSP streams from 1000 service centers.
- PyTorch
- LoRA
- SimPO
- Flask
- FastAPI
- KServe
- MilvusDB
- Celery
- Redis
- NVIDIA DeepStream