MD Sharique works on extracting and modeling information from images, documents, and transaction networks. His contributions include collaborative research on stroke segmentation and co-invention of technology for extracting tables and answering questions about their contents. In work presented with Somya Rai at EXL, he explains how transaction relationships can improve fraud detection while keeping scoring practical to serve.
From medical images to document structure
He studied at IIT Madras from 2015 to 2020, completing a dual degree in Engineering Design with a specialization in Biomedical Engineering. He co-authored the 2019 preprint Parallel Capsule Net for Ischemic Stroke Segmentation, which investigated automated identification of stroke lesions in MRI images. The team combined parallel network pathways and dense connections with a loss function that accounted for both lesion regions and their boundaries. This addressed a central difficulty in segmentation: preserving small, clinically relevant structures within a much larger background.
His May 2022 article on extracting information from unstructured documents treats page layout as information the model should use. Pages become images; a vision model locates tables, headings, paragraphs, and signatures; the application selects relevant regions; and text-processing models extract the required content. Locating an object first narrows what the downstream model must interpret and allows different types of content to receive different treatment.
He is one of eleven named inventors on a patent for multimodal table extraction and semantic search, assigned to ExlService Holdings and granted in March 2025. The collaborative invention connects visual extraction to querying: it detects table cells, converts their contents into machine-readable structures, identifies relevant rows and columns, and answers natural-language questions. Its examples include recognizing equivalent medical concepts expressed through different domain-specific labels.
Preparing information and choosing what to optimize
His writing extends beyond extraction to decisions about model development and deployment:
Data preparation before retraining: In his January 2024 article on knowledge graphs and language models, he questions whether every new business problem needs another annotation and training cycle. He advocates considering pretrained models and investing in data preparation to shorten development. The decision is whether the problem requires a new model or better-organized information for an existing one.
Inference techniques matched to constraints: His 2026 Taming the Transformer series covers memory management, batching, quantization, pruning, parallelism, and attention optimization. The concluding installment distinguishes techniques used in production inference stacks from emerging research, helping readers assess both the constraint an optimization addresses and its maturity for deployment.
Fraud detection within an analyst alert budget
In Judge a Transaction by the Company It Keeps, MD Sharique and Somya Rai describe an EXL system that adds graph embeddings to an existing XGBoost fraud classifier. Somya introduces the value of transaction neighborhoods and the constraints of two portfolios; he develops the engineering choices around sampling, graph construction, explanation, and serving. Their objective is to detect more fraud at a matched alert rate, so improvements fit within analysts’ capacity to investigate.
The graph must preserve useful relationships without exposing the model to future information. Selected users retain their transaction histories during sampling, rather than leaving disconnected fragments of their neighborhoods. Separate monthly graphs, earlier timestamps, and preprocessing fitted only on training data constrain leakage. GraphSAGE combines mean, maximum, and standard deviation when aggregating neighborhood information: an unusual neighbor can remain visible even when averaging would obscure it.
The serving design moves expensive graph computation away from individual scoring requests. Cached representations support a small online forward pass; a shift from transaction-node embeddings to user and merchant embeddings reduces the reported cache footprint from roughly 60 GB to 2–3 GB. Edge attribution connects suspicious scores to particular transaction relationships for analyst review.
Together, the presenters report relative detection improvements of 27%–33% over their tabular baseline at matched alert rates. These are results for the portfolios discussed, rather than a universal fraud benchmark. The financial estimate is projected annualized incremental fraud capture, not realized savings.
Somya Rai and MD Sharique explain how EXL adds graph embeddings to XGBoost for fraud detection—and why preserving transaction neighborhoods, preventing leakage and keeping graph work off the request path matter as much as the model.
Graph embeddings add neighborhood information to XGBoost. The reported 27–33% relative detection gains compare the combined system with a tuned tabular baseline at matched alert rates.
Sampling must preserve useful connectivity: retain selected users’ histories, keep legitimate transactions from fraud-associated users as hard negatives, and balance training batches downstream.
Mean, maximum and standard deviation preserve different neighborhood signals; maximum pooling keeps a single suspicious neighbor from disappearing into an average.
Caching removes graph-sized work from each request, but memory and freshness remain constraints. User-and-merchant embeddings reduce the stated cache from 60 GB to 2–3 GB; the final reported latency is 3.1 milliseconds at P99.
▶ Watch ↗