Swapnil Shinde
Ashburn
Summary
Result oriented professional with 14+ years of experience in AI research and delivering scalable, intelligent & responsible AI systems for fintech and advertising industries. Passionate about AI research, system engineering, and optimizations.
Overview
18
18
years of professional experience
1
1
Certification
Work History
Distinguished Engineer (Director level)
Capital One
02.2024 - Current
- Architected a SOTA recommender backbone model powering bespoke personalization problems, achieving 8,000 TPS at <250ms latency and driving hundreds of millions in incremental revenue. The RecSys platform is designed with responsible AI as a cornerstone, considering critical financial decision-making and governance.
- Delivered scalable STELP system as a multi-agentic solution for Capital One's AI initiatives covering chatbots, intelligence summarization, and AI-based software development workflows.
- Developed STELP (Secure Transpilation & Execution of LLM-Generated Programs) to convert unsafe code into secure applications.
- Created AI-based content generation and validation system, ensuring AI safety and ethical standards in application.
- Conducted AI research prioritizing safety and responsible practices in multi-agentic systems, contributing to ethical AI development.
- Applied AI research and engineering leadership on recommendation systems and multi-agentic frameworks.
- Contributed to billions of AI-driven personalized content deliveries through advanced research and engineering efforts.
- Published papers at academic and industry conferences showcasing innovative findings.
Senior Manager (Sr. Lead), Data & ML Engineering
Capital One
02.2020 - 01.2024
- Leading Data and ML engineering team focused on building an experimentation and personalization platform for Capital One.
- Developed an overlapping experimentation framework, metric computation, statistical engine & insights systems for enterprise experimentation platform.
- Developed a cross channel recommendation platform to personalize experiences (marketing, servicing, feed etc.) for digital customers. This platform is composed of several objective specific ML systems and a global optimizer to select optimal experiences to be served to a customer in real time (< 400ms). This platform provides scalable and extensible ML engineering design empowering every phase of ML lifecycle - Real-time and batch feature engineering, Model training/tuning, Serving, Monitoring and Explainability.
- Strategized & developed roadmap and vision of Capital One's experimentation and unified personalization with product and data science leaders.
- Led contributor on design and roadmap of ML engineering standards, platforms, tools to democratize and scale ML in Capital One. e.g., Feature platform, Batch and streaming feature compute, Realtime model serving, ethical AI etc.
Lead software engineer, Innovation
Comscore Inc.
12.2018 - 01.2020
- As part of Comscore's Innovation team, I focused on research and development of innovative methodologies, algorithms and designing end to end machine learning data pipelines.
- Panel fusion algorithm - Research and development of a scalable fusion algorithm to combine two disparate massive panel (sample) datasets using distributed relaxed min cost flow algorithm in Spark. Developed the same algorithm into a generic service to solve combinatorial assignment problems in Comscore. A research paper was published in the 57th Annual Allerton conference.
- Scalable raking algorithm - Developed a scalable version of the widely used raking algorithm to generate projection weights for sample datasets in Spark. This iterative sequential stratification algorithm is able to generate weights for a massive sample of 30+ Million entities within 20 minutes.
- Display images ad detection system - Designed an end-to-end deep learning system to detect ads in display images served on the internet. Implemented a neural network using Tensorflow to predict if a given web image is a potential ad.
- Demographic assignment model algorithm - Contributing to designing a demographic prediction model using a neural network to predict a demographic profile of an unknown internet cookie or device ID. It involves distributed training of neural networks using Spark.
Software engineering manager
Comscore Inc.
07.2015 - 12.2018
- Led core processing design team to develop next-generation architecture and data systems on more than 18 petabytes of information rich datasets.
- Methodology as a service platform - Inventor and primary author of a self-serving platform using Spark to provide the feature engineering, methodology and algorithmic components as building blocks. This next generation architecture successfully transformed Comscore's research and product development practices becoming more efficient, disciplined and incrementally iterative.
- Census-based panel inference system - Developed end to end data architecture and scalable algorithms to infer demographic profile and media consumption behavior (including co-viewing) from anonymized digital, TV & OTT traffic.
- Built and led a successful team of engineers to solve cross-platform audience measurement problems with higher engineering standards.
Senior software engineer
Comscore Inc.
05.2011 - 07.2015
- Played various roles starting as an intern in the core processing team building innovative cross-platform solutions on the massive scale of internet and television datasets.
Software engineer
IBM India
06.2008 - 07.2010
Education
MS - Computer Science
George Mason University
Fairfax, VA05-2012
Bachelor of Engineering - Electronics
Walchand College of Engineering
Sangli, India05-2008
Skills
- Artificial Intelligence (RecSys, LLMs)
- Reinforcement learning
- AI red teaming
- Multi-agentic systems
- PyTorch
- Python
- Spark
Certification
- DeepLearning.AI TensorFlow Developer Specialization
- IBM Deep Neural Networks with PyTorch
- DeepLearning.AI Deep Learning Specialization
Publications
- Adaptive Instruction Composition for Automated LLM Red-Teaming. ACL 2026 Main Conference.
- ART: Adaptive Reasoning Trees for Explainable Claim Verification. EACL 2026.
- Building safe GenAI applications: An end-to-end overview of red teaming for large language models. Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025), 335-350
- STELP: Secure Transpilation and Execution of LLM-Generated Programs. Arxiv preprint 2025
- A Multi-Stage Workflow for the Review of Marketing Content with Reasoning Large Language Models. arXiv preprint 2025
- Scalable Panel Fusion Using Distributed Min Cost Flow, 2019 57th Annual Allerton Conference on Communication, Control, and Computing (Allerton), Allerton Park and Retreat Center, Monticello, IL, USA, 09/24/19, https://proceedings.allerton.csl.illinois.edu/media/files/0048.pdf
- Time Travel is Real - Building Offline Evaluation Framework, Medium blog, 02/01/21, https://bit.ly/3Cc0lGN
- ML Engineering at Capital One, Blog, 10/01/21, https://bit.ly/3pE1Ny9
- Sequential Feature Encoding/Decoding Optimization Blog
Patents
- SYSTEMS AND METHODS FOR EXECUTING QUERIES ON A BITMAP INDEX. Patent number - 11,966,936; Application number - 17/317,722
- SYSTEMS AND METHODS FOR EXECUTING QUERIES ON A BITMAP INDEX. Patent number - 12,340,384; Application number - 18/607,735
- SYSTEMS AND METHODS FOR DETERMINING TARGET POPULATIONS FOR STATISTICAL EXPERIMENTS. Application number - 17/870,881
- Iterative Communication Plan Generation Using Large Language Models. Application number - 18/820,532
- SYSTEMS AND METHODS FOR PROCESSING DATA OBJECTS IN A PROTECTED NETWORK TO BLOCK TRANSMISSION OF A SUBSET OF THE DATA OBJECTS THAT ARE SUBJECT TO NETWORK VIOLATIONS. Application number - 19/212,565
- SYSTEMS AND METHODS FOR IMPROVING SECURITY BY IDENTIFYING ATTACK VECTORS. Application number - 19/212,576
- SYSTEMS AND METHODS FOR GENERATING SECURED DATA SAMPLES FOR EXECUTING SECURED COMPUTER NETWORK FUNCTIONS. Application number - 19/212,592
- STATEFUL, MULTI-STAGE ACTION PLAN GENERATION FOR A LARGE LANGUAGE MODEL-BASED CREDIT CARD CUSTOMER CONVERSATIONAL ASSISTANT. Application number - 19/387,577
- SYSTEMS AND METHODS FOR ACCESSING EXTERNAL TOOLS DURING EXECUTION OF LARGE LANGUAGE MODELS. Application number - 19/387,580
- USING CLUSTERING TO REDUCE LARGE LANGUAGE MODEL OVERHEAD. Application number - 19/424,994
- VALIDATING CONTENT AGAINST RULES USING A LARGE LANGUAGE MODEL AND COMPUTER VISION. Application number - 19/425,019
- STELP: Secure Transpilation and Execution of LLM-Generated Program. Application number - 63/774,707
- ADAPTIVE INSTRUCTION COMPOSITION FOR AUTOMATED LLM RED TEAMING. Application number - 63/804,554
- ORCHESTRATED HYBRID RAG: A SCALABLE ARCHITECTURE FOR HIGH-THROUGHPUT MARKETING LEAD ENRICHMENT. Application number - 63/977,796
Academic Activities
- Reviewer on EMNLP industry track 2025, ICML 2026, NeurIPS 2026, etc.
- Academic and professions memberships.
Timeline
Distinguished Engineer (Director level)
Capital One
02.2024 - Current
Senior Manager (Sr. Lead), Data & ML Engineering
Capital One
02.2020 - 01.2024
Lead software engineer, Innovation
Comscore Inc.
12.2018 - 01.2020
Software engineering manager
Comscore Inc.
07.2015 - 12.2018
Senior software engineer
Comscore Inc.
05.2011 - 07.2015
Software engineer
IBM India
06.2008 - 07.2010
MS - Computer Science
George Mason University
Bachelor of Engineering - Electronics
Walchand College of Engineering