← Return to OverviewAcademic Research Focus

Research Vectors & Empirical Methodology

My research concentrates on developing grounded, verifiable machine learning systems operating under structural constraints—whether bridging symbolic knowledge graphs with neural retrieval or optimizing quantized inference on microcontrollers.

Vector 01

LLMs & Retrieval-Augmented Generation (RAG)

Standard generative models frequently suffer from hallucinations and lack structural provenance when queried on specialized corpora. My work focuses on constructing structured retrieval pipelines that ground language model reasoning within verifiable data substrates.

Knowledge Graph Integration

Investigating structured retrieval where entity-relationship knowledge graphs provide topological constraints and explicit reasoning paths for retrieval-augmented generation.

Multimodal & Domain Retrieval

Extending structured retrieval to multimodal inputs and specialized financial and academic documentation where tabular, textual, and temporal formats intersect.

Vector 02

Financial AI & Econometric Network Dynamics

Traditional market modeling relies heavily on linear Pearson correlation, systematically underestimating systemic tail risks and cross-sector phase shifts. By deploying non-parametric rank statistics, time-lagged cross-correlation, and graph reduction via Minimum Spanning Trees (MST), non-linear co-movements and rotational lead–lag phenomena can be modeled robustly.

The Linear Risk Trap & AvgDiff

Demonstrated an average divergence of ~0.40 between linear and non-linear correlation structures across ten major Indian sectors, proving that linear risk frameworks overlook systemic risk blocks.

Lead–Lag Dependency Isolation

Utilized time-lagged cross-correlation to detect directional flow signals—such as the Pharma sector leading the Auto sector by an empirical window of -75 days.

Vector 03

Edge AI & TinyML Physical Intelligence

Physical AI systems cannot depend on continuous cloud connectivity or unlimited compute budgets. My work in TinyML addresses real-time neural inference directly on microcontrollers through motion feature engineering, model pruning, and post-training quantization.

On-Device Quantization

Quantizing neural networks into 8-bit integer formats executable on low-power microcontrollers such as the Arduino UNO Q without catastrophic degradation in precision.

High-Frequency Sensor Processing

Processing raw multi-axis accelerometer and gyroscope time-series streams in real-time to detect anomalous physical events such as sudden falls.

Vector 04

Machine Learning Foundations & Anomaly Detection

Developing robust anomaly detection systems that integrate statistical rules with machine learning and large language model reasoning signals, backed by human-in-the-loop review mechanisms for high-stakes decision reliability.