Somya Rai

Somya Rai builds AI systems for insurance and financial risk, from underwriting assistants that coordinate document checks and human review to fraud models that use relationships between transactions. A vice president and AI architect at EXL, she also co-authored Token Economics, which examines how model selection and inference design affect operating costs.

Insurance workflows and agent design

Somya’s background spans banking, insurance, and risk. She studied at the Indian Institute of Management Bangalore in 2019–2020. By 2023, she was EXL’s principal AI solution architect, applying machine learning, deep learning, natural-language processing, and generative AI to insurance problems. In 2025, she described her role as principal AI engineer.

An underwriting assistant developed by her team shows what those applications require. Commercial insurance submissions arrive as emails and attachments containing documents, images, and location information. Assessing a business can involve checking sanctions databases, identifying its industry, examining nearby hazards, and comparing its risks with previously underwritten businesses. The team divided that work among agents while retaining human review at consequential steps.

The initial design exposed a specific failure: combining validation and checking in one agent gave it too many tools and produced unreliable results. The team separated those responsibilities. It also used LangGraph to preserve workflow state, so an unavailable reviewer or expired session would not force the process to restart. In her account of the underwriting assistant, Somya reported that the deployed system reduced an insurer’s underwriting turnaround from weeks to two days. The result belonged to the team; the design decisions explain how it accommodated the insurer’s working practices.

Relationships as fraud signals

In Judge a Transaction by the Company It Keeps, Somya and MD Sharique presented an EXL fraud detection system that adds graph embeddings to an existing XGBoost classifier. Somya introduced the central problem and the constraints of two transaction portfolios: a transaction can appear harmless on its own while its connections to recent fraud make it suspicious. Adding more features about the individual transaction can miss that neighborhood signal.

The objective was to detect more fraud within a fixed analyst alert budget. Their joint presentation reported relative detection improvements of 27% to 33% over the tabular baseline at matched alert rates. This comparison matters because sending analysts more alerts would otherwise make an apparent improvement difficult to interpret. The financial figure discussed was projected annualized incremental fraud capture, rather than realized savings.

MD Sharique presented the detailed sampling, graph construction, explanation, and serving methods. Those engineering details support the joint result without establishing Somya’s individual authorship of each component.

The economics of running models

Somya’s technical writing examines several decisions that determine whether a model can serve its intended workload:

• GPU utilization and batching: Her guide to LLM inference distinguishes computational capacity from memory bandwidth. A GPU can have arithmetic capacity sitting idle while a model waits for data. Batching requests can improve throughput, but waiting to assemble a batch adds delay: document-processing queues can tolerate waits that interactive applications may not. She treats hardware calculations as a starting point to test with benchmarks, including communication overhead.

• Task-aware model routing: In Token Economics, co-authored with EXL chief AI officer Andy Logani, Somya argues for allocating requests across a portfolio of models. Routine extraction or classification can go to a smaller model, with harder or uncertain cases escalated. The paper also examines speculative decoding, in which a smaller model proposes tokens for a larger model to verify. Both approaches make cost and response time architectural decisions.

• Generative recommendation: Her writing on recommender systems examines the shift from retrieving and ranking items through embedding similarity toward predicting subsequent interactions from an ordered history. She connects that change to infrastructure constraints: interaction histories change continuously, catalogs have sparse long tails, and the working data can exceed accelerator memory.

Neural approximations for insurance reserves

Somya co-authored the September 2026 research paper Insurance Reserve Intelligence Platform with Anugya A, Saket Mohanty, and Abhilash Timmapur. The project combines a classical actuarial solver with a neural approximation for term-life insurance reserves—the funds an insurer holds for future obligations.

Its training objective penalizes violations of the governing actuarial equation and incorporates additional insurance constraints. The aim is to accelerate repeated scenario calculations while retaining actuarial structure in the model. Evaluated using synthetic policies, the work remains a prototype with weaknesses in generalization and some expected reserve behavior. Those limitations qualify its promise: faster approximation still has to reproduce the behavior that makes reserve calculations useful.