Research & Scientific Inquiries
Bridging rigorous mathematical foundations with modern computational architectures to address heritage preservation and financial market microstructure.
Primary Research Interests
Document Epigraphy & Vision Transformers
Degraded historical character segmentation, morphological noise reduction, and transformer-based OCR for Indic palm leaves.
Market Microstructure & Order Book Dynamics
Deterministic sub-microsecond limit order execution, memory arena pools, and stochastic Hawkes processes.
Computational Finance & Stochastic Volatility
Antithetic Monte Carlo option pricing, Heston stochastic volatility surfaces, and Greek sensitivity risk metrics.
Parameter-Efficient LLM Fine-Tuning (PEFT / LoRA)
Low-rank adaptation architectures for domain-specific algorithmic problem solving and technical evaluation.
Active & Completed Research Projects
Indic Palm Leaf Manuscript Analysis & Neural Digitization
Ancient Indic palm leaf manuscripts suffer from ink fading, physical decay, and complex character overlap. This project develops a deep learning pipeline combining advanced image enhancement, character segmentation, and convolutional neural architectures to transcribe and classify degraded historical scripts accurately.
Methodology
Employed custom contrast-limited adaptive histogram equalization (CLAHE) with morphological filtering for foreground script isolation, followed by a fine-tuned ResNet and Vision Transformer backbone for multi-class glyph classification.
Dataset
Curated dataset of high-resolution Indic manuscript scans provided in collaboration with IIT Tirupati researchers.
Architecture & Models
ResNet-50, Vision Transformer (ViT), U-Net Segmentation
Outcome: Achieved 94.6% character recognition accuracy on degraded palm leaf specimens, significantly outperforming baseline OCR models on low-contrast historical documents.
High-Performance Limit Order Book & Low-Latency Matching Engine
Modern financial markets require deterministic, sub-microsecond order execution. This research implements a cache-optimized limit order book matching engine in C++ and Python, evaluating price-time priority (FIFO) queues under synthetic high-frequency order flows.
Methodology
Utilized contiguous memory arrays, custom memory pools, and lock-free ring buffers to minimize cache misses and latency spikes during high-burst market volume simulations.
Dataset
Synthetic high-frequency market order flow and historical tick data.
Architecture & Models
Stochastic Hawkes Process, Poisson Order Arrival Simulator
Outcome: Sustained over 2.4 million order matches per second with mean latency under 420 nanoseconds in optimized C++ benchmark environments.
Monte Carlo Simulations for Exotic Option Pricing & Risk Modeling
Evaluated European and Asian option pricing under Black-Scholes and Heston stochastic volatility models using variance-reduced Monte Carlo simulation techniques including antithetic variates and control variates.
Methodology
Simulated 100,000+ stochastic price trajectories via Geometric Brownian Motion and Milstein discretization, analyzing convergence rates and Greek sensitivities (Delta, Gamma, Vega).
Dataset
Simulated market paths and historical S&P 500 equity options volatility surfaces.
Architecture & Models
Black-Scholes-Merton, Heston Stochastic Volatility Model
Outcome: Reduced pricing variance by 41% with antithetic sampling, delivering accurate numerical estimates for non-linear exotic payoff structures.
Parameter-Efficient Fine-Tuning (LoRA) for Specialized Technical Agents
Investigated low-rank adaptation (LoRA) and quantized LoRA (QLoRA) parameter efficiency for domain adaptation of 7B-parameter open-weights models (Llama and Mistral) for scientific problem solving and interview guidance.
Methodology
Applied rank-8 low-rank decomposition to attention projection weights, minimizing trainable parameters to under 0.2% while retaining generative quality across technical question datasets.
Dataset
Domain-curated STEM benchmarks and algorithmic coding interviews.
Architecture & Models
Mistral-7B, Llama-2/3, vLLM High-Throughput Inference Engine
Outcome: Trained models achieved 18% improvement on domain-specific mathematical and coding evaluations while training in under 1/4th the GPU memory of full fine-tuning.
Interested in Research Collaboration?
I welcome research inquiries, graduate lab discussions, and co-authorship proposals in deep learning vision systems and quantitative financial architectures.