Hardik Srivastava

About

I am a Masters in Data Science student at the University of Washington. I am associated with the Washington AI Lab at the Paul G. Allen School of Computer Science & Engineering, advised by Prof. Zaid Harchaoui. My current research focuses on post-training optimization of LLMs, with a core emphasis on Alignment and Privacy. I am working on the mechanics of style-aware text rewriting and stylistic perturbation to enable authorship obfuscation. To achieve this, I work on building multi-agent systems and train parameter-efficient ensembles (LoRAs) that capture distinct stylistic features via attribution-guided modeling. Applying custom constrained decoding algorithms at inference enables me to dynamically steer LLM generation and meet strict privacy-utility trade-offs. Beyond model alignment, I am passionate about inference optimization, designing scalable, low-latency agentic architectures that maintain robust performance at enterprise scale.

Prior to UW, I spent over two years as an Applied Scientist at JPMorgan Chase within the JPMorgan Payments - AI Trust and Safety team, where I led the applied research and deployment of Gradient Boosting, multimodal Transformers, and Self-Supervised Graph ML models. Focusing on large-scale adversarial modeling and anomaly detection, I engineered solutions that captured complex behavioral anomalies in massive networks, preventing over $220M in losses annually. My work spanned architecting Named Entity Recognition pipelines, Entity Resolution models for sanctions screening, and Learning-to-Rank information retrieval systems, leveraging tools like Spark, MLflow, Kubernetes, and AWS to productionize these pipelines at enterprise scale.

My foundational research experience at Samsung Research includes shipping an optimized NLU Intent Routing system into production for Samsung Bixby assistant along with deploying latency-aware INT8 quantization for word-sense disambiguation models, matching FP32 performance via Quantization-Aware Training. I also conducted research in Human-Centered AI and multimodal NLP at the Accessible Computing Lab (ACT) at McGill University.

I am passionate about building AI systems that are rigorous, reproducible, and grounded in theory, while being explicitly engineered for deployment at enterprise scale.

Let's connect if your team is building something at the intersection of learning, reasoning, and trustworthy AI.

Research Interests

My core research interests lie at the intersection of Natural Language Processing, Generative AI, Graph Representation Learning, Human-AI Interaction, Causality, and Trustworthy ML. I focus on designing language and multimodal systems that are interactive, adaptable and explainable to diverse user needs, and capable of operating reliably in high-stakes environments.

• Efficient Transformer Architectures for low-resource language generation
• Behavioral AI and Human-Computer Interaction for Cognitive Modeling
• AI Alignment, Privacy-Preserving ML & Explainability
• Human-in-the-loop NLP and user-adaptive systems

News

11/2025
Secured 3rd place at the UW x Databricks Hackathon 2025 held at Seattle, WA. We built KeyCognition - an ML system that predicts cognitive load from keystroke dynamics by using fine-grained keystroke features.
09/2025
I am starting a new position as a Graduate Research Assistant at UW under Prof. Zaid Harchaoui
08/2025
I have joined the University of Washington at Seattle, WA as a Masters in Data Science student
09/2024
Secured the 1st Place at JPMC ScaleUP Challenge 2024 for Protection Group Code Prediction for Classified Data Columns using BERT-based Active Learning
08/2024
Delivered a session on how we leverage Hybrid Recommendation models for minimizing risk in my Algo-Trading Platform based solution at JPMorgan's Innovation Week 2024
11/2023
Cleared the AWS Cloud Practitioner and AWS Machine Learning Speciality certifications
09/2023
Our paper Multi-Modal Sentiment Analysis Using Text and Audio for Customer Support Centers has been accepted by ICACTCE 2023.
06/2023
I am starting a new position as Applied Scientist at JPMorgan Chase.
04/2023
My Undergraduate thesis Decision Support Complaint Prioritization System using a Statistical Multi-Method Algorithmic approach has been accepted and the preprint is availble
08/2022
Our paper Neural Text Style Transfer with Custom Language Styles for Personalized Communication Systems has been accepted to ICKECS 2023.
10/2022
We have won the 1st place of SRM Hack2Leap Hackathon 2022.
08/2022
We have won the 1st place of Smart India Hackathon 2022.
06/2022
I started a new position as a Research Intern at ACT Lab, McGill University via the Mitacs Globalink Research Internship to extend my work on Human-Computer Interaction and AAC Systems.
05/2021
Our proposal for a Real-Time Video Stablization and Deblurring software toolkit has been accepted with a grant of $21500 by Defence Research and Development Organization, Govt. of India.
05/2021
Our paper Using NLP Techniques for Enhancing Augmentative and Alternative Communication Applications has been accepted by IJREAM
01/2021
I started a new position as an Undergraduate Research Assistant under Dr. T Rajalakshmi at the Biomedical Engineering department, SRM.

Publications

• Selected
• Conferences
• Journals
• Thesis
• Preprints
• All

Multi-Modal Sentiment Analysis Using Text and Audio for Customer Support Centers

Hardik Srivastava, Sneha Sunil, K. Shantha Kumari, P. Kanmani

ICACTCE: International Conference on Advances in Communication Technology and Computer Engineering (Springer Nature) (2023)

Proposed CM-BERT, a novel multimodal sentiment analysis model that fuses textual and audio features to robustly capture customer sentiment

Paper Code

Decision Support Complaint Prioritization System using a Statistical Multi-Method Algorithmic approach

Hardik Srivastava, Mayank Jha, T. Karthick

Undergraduate Thesis (2023)

Proposed a statistical multi-method framework combining rule-based scoring and algorithmic ranking to integrate diverse decision signals into a coherent prioritization score for efficient re-ranking.

Paper Code

Neural Text Style Transfer with Custom Language Styles for Personalized Communication Systems

Hardik Srivastava, Sneha Sunil, K. Shantha Kumari

ICKECS'22: International Conference on Knowledge Engineering and Communication Systems (IEEE) (2022)

Introduced StyleLM, a style-conditioned neural text transfer model enabling fine-grained control over language style for personalized communication.

Paper Code

Automatic Screening and Staging of Multi-Stage Diabetic Retinopathy using Deep Learning techniques

Hardik Srivastava, T. Rajalakshmi

Preprint (2022)

Proposed a deep learning framework for automated detection and staging of multi-stage diabetic retinopathy.

Paper

Using NLP Techniques for Enhancing Augmentative and Alternative Communication Applications

Hardik Srivastava

IJREAM: International Journal for Research in Engineering Application & Management (UGC) (2021)

NLP-based framework to enhance AAC applications, enabling dynamic vocabulary expansion and more expressive sentence generation.

Paper

Multi-Modal Sentiment Analysis Using Text and Audio for Customer Support Centers

Hardik Srivastava, Sneha Sunil, K. Shantha Kumari, P. Kanmani

ICACTCE: International Conference on Advances in Communication Technology and Computer Engineering (Springer Nature) (2023)

Proposed CM-BERT, a novel multimodal sentiment analysis model that fuses textual and audio features to robustly capture customer sentiment

Paper Code

Neural Text Style Transfer with Custom Language Styles for Personalized Communication Systems

Hardik Srivastava, Sneha Sunil, K. Shantha Kumari

ICKECS'22: International Conference on Knowledge Engineering and Communication Systems (IEEE) (2022)

Introduced StyleLM, a style-conditioned neural text transfer model enabling fine-grained control over language style for personalized communication.

Paper Code

Using NLP Techniques for Enhancing Augmentative and Alternative Communication Applications

Hardik Srivastava

IJREAM: International Journal for Research in Engineering Application & Management (UGC) (2021)

NLP-based framework to enhance AAC applications, enabling dynamic vocabulary expansion and more expressive sentence generation.

Paper

Decision Support Complaint Prioritization System using a Statistical Multi-Method Algorithmic approach

Hardik Srivastava, Mayank Jha, T. Karthick

Undergraduate Thesis (2023)

Proposed a statistical multi-method framework combining rule-based scoring and algorithmic ranking to integrate diverse decision signals into a coherent prioritization score for efficient re-ranking.

Paper Code

Automatic Screening and Staging of Multi-Stage Diabetic Retinopathy using Deep Learning techniques

Hardik Srivastava, T. Rajalakshmi

Preprint (2022)

Proposed a deep learning framework for automated detection and staging of multi-stage diabetic retinopathy.

Preprint

Decision Support Complaint Prioritization System using a Statistical Multi-Method Algorithmic approach

Hardik Srivastava, Mayank Jha, T. Karthick

Undergraduate Thesis (2023)

Proposed a statistical multi-method framework combining rule-based scoring and algorithmic ranking to integrate diverse decision signals into a coherent prioritization score for efficient re-ranking.

Paper Code

Multi-Modal Sentiment Analysis Using Text and Audio for Customer Support Centers

Hardik Srivastava, Sneha Sunil, K. Shantha Kumari, P. Kanmani

ICACTCE: International Conference on Advances in Communication Technology and Computer Engineering (Springer Nature) (2023)

Proposed CM-BERT, a novel multimodal sentiment analysis model that fuses textual and audio features to robustly capture customer sentiment

Paper Code

Automatic Screening and Staging of Multi-Stage Diabetic Retinopathy using Deep Learning techniques

Hardik Srivastava, T. Rajalakshmi

Preprint (2022)

Proposed a deep learning framework for automated detection and staging of multi-stage diabetic retinopathy.

Paper

Neural Text Style Transfer with Custom Language Styles for Personalized Communication Systems

Hardik Srivastava, Sneha Sunil, K. Shantha Kumari

ICKECS'22: International Conference on Knowledge Engineering and Communication Systems (IEEE) (2022)

Introduced StyleLM, a style-conditioned neural text transfer model enabling fine-grained control over language style for personalized communication.

Paper Code

Using NLP Techniques for Enhancing Augmentative and Alternative Communication Applications

Hardik Srivastava

IJREAM: International Journal for Research in Engineering Application & Management (UGC) (2021)

NLP-based framework to enhance AAC applications, enabling dynamic vocabulary expansion and more expressive sentence generation.

Paper

Projects

Recursive Episodic Memory (REM) RAG
Engineered a memory compression framework for long-context LLM agents. Utilizing FAISS-based dense retrieval, hierarchical semantic indexing, and adaptive replay mechanisms, this architecture mitigates context decay and enables persistent, multi-session agentic reasoning across complex computational pipelines.

PokerFace: RL for Strategic Deception in LLMs
Trained LLMs for strategic deception and negotiation in a Market for Lemons env using PPO. Engineered a two-stage reasoning pipeline to maximize strategic rewards without leaking hidden states to opponents. Scaled training via QLoRA for parameter-efficient fine-tuning on instruct models.

Ad-Astra: Agentic Ad Intelligence Engine
Built a dynamic ad generation and placement engine for X (Twitter), using Grok 4.1 Reasoning models. Utilized behavioral scroll telemetry and temporal interaction signals to optimize in-feed ad placement improving relevance and engagement.

KeyCognition: Predicting Cognitive Load via Keystrokes
Behavioral AI platform that predicts user cognitive load in real-time using temporal sequence modeling. Engineered a React app to extract micro-temporal typing anomalies like dwell time & flight time using 136M keystrokes.

Auto-Explainability & Evaluation of LLMs
Transformer explainability framework designed to demystify language model generation. Utilizes Integrated Gradients to map and visualize token-level prediction attributions enabling deep interpretability.

No-Code Algo-Trading Portfolio Manager
Domain-Specific Language (DSL) framework powering a no-code trading platform. Leverages hybrid recommendation models to dynamically compile drag-and-drop user inputs into optimized algorithms, thereby minimizing risk and expected shortfall.

Articles

MLOps 101 — Machine Learning Workflow Orchestration using TensorFlow Extended
Deploying Machine Learning workloads efficiently in Production using TFX

A/B Testing in Data Science
Learning how to let the customer do the talking, without actually talking.