Projects
Research & Publications
Peer-reviewed work in NLP, multi-agent reinforcement learning, and robotics.
-
EACL 2026 (main conference) · October 2025A new benchmark for granular evaluation of LLM instruction compliance abilities, enabling fine-grained analysis of where language models fail to follow constraints.
-
EACL 2026 (main conference) · October 2025An evaluation-driven multi-agentic workflow for prompt instruction optimization, capable of automatically improving the constraint-following ability of LLMs.
-
IEEE Transactions on Robotics · March 2024Multiagent reinforcement learning via rollout and policy iteration for partially observable Markov decision processes (POMDPs), demonstrated on multi-robot coordination tasks.
-
COLING 2022 · October 2022A data-efficient training approach to Aspect Based Sentiment Analysis using unsupervised data augmentation, reducing dependence on labelled data in low-resource settings.
-
Conference on Robot Learning (CoRL) · November 2020Multiagent rollout and policy iteration for POMDP, with application to multi-robot repair problems in complex, partially observable environments.
-
IEEE Robotics and Automation Letters · July 2020Partitioned rollout and policy iteration for autonomous sequential repair problems in partially observable environments, published in IEEE RA-L.
-
IEEE AMIAMS Conference · May 2016AutoCAD plugin using a genetic algorithm to generate architectural space layouts satisfying geometric and topological constraints, giving architects multiple design options automatically.
Industry Projects
Selected production systems and applied research from 7+ years of industry work.
-
DFS Risk Model IntegrationCapital One · 2025 – PresentLeading the integration of Discover Financial Services Customer Management Risk Models (valuation $3B+) with Capital One's infrastructure.
-
LLM Instruction Benchmark & Prompt OptimizerCapital One · 2024 – 2025Invented a new benchmark to evaluate LLM instruction-following and developed a multi-agentic prompt optimization system to automatically improve compliance. Work published at EACL 2026.
-
Conversational Servicing AgentCapital One · 2023 – 2024Built a tool-augmented conversational agent using ReACT, RAG, and LangChain, with a fine-tuned Mistral-7B backend (SFT + PEFT: LoRA, DoRA, GaLore) for reliable tool calling.
-
Self-Play Multi-Agent Fraud SimulationCapital One · 2024Designed a LangGraph-based self-play game between Agent, Customer, and Fraudster personas using ReACT and Reflexion for short- and long-term planning, surfacing novel attack patterns and countermeasures.
-
Row Encoder Transformer for Fraud DetectionCapital One · 2023Designed a row-level encoder transformer for tabular fraud detection that outperformed the GBM baseline by 5% at scale. Also explored time-series transformers (TimesFM, iTransformer) for financial forecasting.
-
Call Intent Classification at ScaleCapital One · 2021 – 2023Built a large multi-label BERT-based classifier (350+ categories) on call transcripts to identify why customers contact Capital One, processing 300K+ calls per day in production.
-
Deep Learning Video Recommendation EngineRoposo (Relevant E Solutions) · 2018 – 2019Replaced a traditional ALS recommendation system with a deep learning collaborative filtering model, lifting video views by 2% across 14M users. Validated via A/B testing; trained on distributed TensorFlow on AWS.
-
Contact-Center NLP PipelineAmeyo (Drishti Soft Solutions) · 2016 – 2018Built an end-to-end NLP pipeline for sentiment and emotion analysis of inbound contact-center emails using an LSTM on TensorFlow (F1 ~0.86). Also developed a BTTC (best time to contact) model that improved outbound call connectivity by 2% at HDFC Bank.