Hey there!
I'm Arnav

Speculative Actions: A Lossless Framework for Faster Agentic Systems

International Conference on Learning Representations (ICLR)

This paper tackles the latency bottlenecks in AI agent systems, where each action typically requires a slow, sequential API call, making even simple tasks like playing a full game of chess take hours. The authors introduce speculative actions, a lossless framework in which faster “guessing” models predict likely next actions so that multiple steps can be executed in parallel without degrading correctness. Evaluated across gaming, e-commerce, web search, and a lossy extension for operating systems, the method achieves up to 55% accuracy in next-action prediction, yielding substantial reductions in end-to-end latency. The study further shows that stronger guessing models, top-K action proposals, multi-step speculation, and uncertainty-aware optimization can amplify these gains, pointing toward more practical, low-latency deployment of complex agentic systems.

Use of spatio-temporal features for earthquake forecasting of imbalanced data.

(IEEE) International Conference on
Intelligent Innovations in Engineering and Technology (ICIIET)

This paper addresses the challenge of predicting large earthquakes from imbalanced seismic datasets, where high-magnitude events are rare compared to smaller ones. I propose transforming time-series earthquake catalogs into feature-rich datasets by incorporating temporal and geospatial indicators such as fault density. Three machine learning approaches—weighted Support Vector Machines, distance-weighted K-Nearest Neighbors, and weighted Decision Trees—are evaluated across multiple seismic regions including the Himalayas, Central Java, Sumatra, Sulawesi, and Southeast Asia. Results show that distance-weighted KNN outperforms the others in accuracy, precision, and F1 score, demonstrating its robustness against data imbalance. The study highlights the potential of spatio-temporal feature engineering and algorithm-level adjustments for more reliable earthquake forecasting.

Disease Identification in Tomato Leaf using pre-trained ResNet and Deformable Inception

(Springer) 5th International Conference on
Computational Intelligence in Data Science

This paper presents a deep learning approach for detecting tomato leaf diseases by combining ResNet-50 with Inception modules and deformable convolutions. Unlike previous models trained mostly on lab-curated datasets, the proposed architecture is evaluated on both controlled (PlantVillage) and real-world farmland (PlantDoc) images, as well as an augmented dataset to improve robustness. The model achieves state-of-the-art accuracy—99.08% on PlantVillage and 66.06% on PlantDoc, significantly outperforming prior methods. By leveraging skip connections, multi-scale filters, and deformable kernels, the approach enhances disease recognition in realistic agricultural settings, offering a promising step toward practical crop disease monitoring systems.

Forecasting Earthquakes Using Neural Network Models.

(Springer Nature) Disaster Management in Complex Himalayan Terrains
Natural Hazard Management, Methodologies and Policy Implications

This chapter explores the use of artificial neural networks to forecast earthquakes in the Himalayan region, one of the most seismically active zones in the world. Using seismic data from 1980 to 2020, I extracted eight key seismicity indicators based on the Gutenberg Richter law and other empirical relations to capture intrinsic earthquake patterns. A neural network architecture optimized with deep learning techniques achieves 90% accuracy and an F1-score of 0.89, demonstrating its effectiveness in modeling the nonlinear and heterogeneous nature of seismic processes. The study underscores the potential of machine learning in advancing earthquake hazard assessment and guiding disaster preparedness in vulnerable Himalayan communities.

Model Integration and Deployment August 2023 - April 2025

Barclays (Team - Model Implementation Team)

• Contributed to the curation of model ready datasets for multiple teams within Barclays
• Spearheaded the design and implementation of a unified messaging service integrating diverse team services.
• Built and managed infrastructure to integrate and run over 100 quantitative models for the bank

My time here at Barclays has introduced me to the finance industry and how Data Science is used in the banking domain. Here I have learnt how data is provisioned to create model ready datasets, specifically in the financial context. My experiences working with etl processes here have have made me wonder about the possibility of automation in data transformation which would significantly speed up the process and reduce the amount of effort spent. Here I developed a model integration tool which automated the integration and deployment of models on the cloud; to be accessed by my team's microservices based application to run the models. During my time here, I had the privilege of participating in a global Generative AI hackathon, which brought together 2,000 participants from across the organization. Collaborating in this innovative and competitive environment, my team and I developed a solution using a Large Language Decoder Model to address a pressing business challenge in the Anti-Money Laundering area. Judged by senior stakeholders, our solution was well received during the presentation phase, and we secured the first position in the hackathon.

Western Australia Transforming Community Health Jan 2022 - June 2022

Western Austrlia Department of Health

• Analyzed ~19000 attributes for 373 suburbs in the Australian continent for improving community health.
• Implemented heirarchical clutering and PCA based clustering for social determinants of health based attribute correlation
• Obtained a specific suburb from the data for in-depth analysis and evaluation of policy effectiveness.

Fresh off my industry experience at Amazon, I worked on my final semester thesis with the Western Australia Department of Health, a government body which works to improve and protect the health of the community in the Western Australian suburbs near Perth. This experience honed my ability to tackle complex data challenges and develop actionable insights that have a positive impact on communities. It also has made me more sensitive about health related problems affecting the world.

User Action Automation June 2022 - December 2022

Amazon (Team - Selection Monitoring)

• Analyzed web domain data for competitor e-commerce websites.
• Utilized AWS resources like Sagemaker, S3, Stepfunctions to implement baseline models for web domain data.
• Constructed a Reinforcement Learning and Webpage Segmentation based approach for user action automation in the web crawler.

My internship at Amazon provided invaluable exposure to the industry, allowing me to gain firsthand insights into the deployment of large-scale models. Moreover, it presented me with a unique opportunity to delve deeply into the realm of reinforcement learning, particularly in the context of automating web tasks. This experience enabled me to explore the intricacies of applying advanced techniques in real-world scenarios and solidified my understanding of the practical implementation of reinforcement learning algorithms.

Portfolio CIO: AI-Powered Investment Intelligence Jan 2026 - Feb 2026

Columbia University

This project was developed as a self-motivated initiative at Columbia University, aimed at democratizing institutional-grade portfolio analytics for retail investors using generative AI.

• Built a React + FastAPI platform with a dark-themed dashboard using Recharts for interactive portfolio visualization
• Designed a multi-agent architecture with seven specialized AI agents (Performance Analyst, Risk Officer, Macro Strategist, Stress Specialist, Allocation Strategist, Goal Planner, and a synthesizing CIO agent) powered by Anthropic Claude
• Computed five risk-adjusted ratios (Sharpe, Sortino, Calmar, Treynor, Information Ratio), tail-risk analysis, and efficient frontier optimization across 2,000 simulated portfolios
• Implemented Monte Carlo stress testing with 5,000-path simulations and historical crisis backtesting across seven insight modules
• Integrated ElevenLabs for audio report generation, WhiteCircle AI for content safety filtering, and PDF report export with portfolio versioning

This project deepened my understanding of quantitative finance concepts such as portfolio optimization and risk analysis, while also strengthening my skills in building multi-agent AI systems, full-stack development with React and FastAPI, and integrating third-party APIs into production-ready applications.

RapidReach: AI-Powered SDR Agent System Jan 2026 - Feb 2026

Columbia University

This project was built at the Columbia ADI DevFest 2026 hackathon, designed to automate the entire sales development lifecycle using a multi-agent AI system.

• Designed a 14-agent architecture distributed across 5 microservices using Dedalus ADK with multi-model LLM routing across Gemini, Claude Sonnet, and GPT-4.1
• Orchestrated a 7-step pipeline covering lead discovery, business research, proposal generation, AI-powered phone calls, outcome classification, deck generation, and email outreach
• Integrated 12+ tools including Google Maps Places API, Brave Search MCP, ElevenLabs Conversational AI for phone calls, and BigQuery for lead management
• Implemented design patterns such as Agent-as-Tool, Coordinator-Specialists, and Generator-Critic for robust agent orchestration
• Built real-time UI updates via WebSocket callbacks and structured output validation with Pydantic

This project gave me hands-on experience with agentic AI design patterns, multi-model LLM orchestration, and building microservices-based systems. It also strengthened my ability to rapidly prototype complex distributed applications under hackathon constraints.

FX Exponential Smoothing Trading Strategy Aug 2025 - Dec 2025

Columbia University

This project was developed as part of a collaborative course project at Columbia University, focused on building and stress-testing a quantitative trading strategy for foreign exchange markets.

• Developed a dual-exponential-smoothing strategy with tunable alpha (slow trend) and beta (fast momentum) parameters, generating trade signals from spread crossovers
• Built a comprehensive backtesting engine with position-based returns, turnover-dependent transaction costs, equity curves, and drawdown analysis
• Implemented grid search optimization across alpha, beta, and threshold combinations with heatmap visualization for parameter tuning
• Integrated Random Forest classification with SHAP-based interpretability for feature importance analysis as a hybrid ML-signal approach
• Supported multi-asset trading across FX pairs, cryptocurrencies, metals, and equity indices using a Streamlit-based interactive research workspace

This project strengthened my understanding of quantitative finance, time-series analysis, and algorithmic trading. It also gave me practical experience with backtesting methodologies, parameter optimization techniques, and combining traditional signal processing with machine learning for financial applications.

AI Agent Safety June 2025 - Dec 2025

Columbia University

This project was undertaken as part of faculty-supervised research at Columbia University, focused on ensuring AI agents complete tasks safely while minimizing potential harm to users.

• Developed a structured side effect evaluation framework that systematically monitors for risks such as data loss, privacy breaches, operations disruption, financial harm, and security violations
• Evaluated agent safety across multiple application domains including file systems, web browsers, document editing, and databases
• Designed methods to capture initial system states and predict potential harmful actions before they occur
• Combined task success metrics with side effect assessments to ensure agents achieve desired outcomes while operating robustly and responsibly
• Explored mitigation strategies enabling agents to act both effectively and safely in complex interactive environments

This research deepened my understanding of AI safety, responsible AI deployment, and the challenges of building reliable agent systems. It equipped me with skills in designing evaluation frameworks, risk assessment methodologies, and systematic testing of AI systems for real-world robustness.

AI Agents Speculative Actions July 2025 - December 2025

Columbia University

This project was undertaken as part of faculty-supervised research at Columbia University, exploring speculative execution as a lossless framework for reducing latency in AI agent systems. The resulting work was published at ICLR.

• Developed a speculative actions framework that treats every agent step (tool calls, LLM calls, human responses) as API calls that can be predicted and executed in advance
• Designed a DAG-based dependency system that captures real dependencies between actions, ensuring correctness while enabling parallel execution of multiple candidate paths
• Achieved up to 55% accuracy in next-action prediction, yielding substantial reductions in end-to-end latency across gaming, e-commerce, web search, and operating system benchmarks
• Explored top-K action proposals, multi-step speculation, and uncertainty-aware optimization to amplify efficiency gains
• Demonstrated applicability across customer service, developer workflows, and simulation platforms, showcasing the viability of speculative execution in interactive systems

This research strengthened my expertise in AI agent architectures, parallel execution strategies, and performance optimization for LLM-based systems. It also honed my skills in experimental design, benchmarking across diverse domains, and publishing research at top-tier venues.

Identifying Disease Using Machine Learning Jan 2022 - May 2022

BITS Pilani

This project was undertaken as part of research at BITS Pilani, applying machine learning techniques to genetic data for disease prediction.

• Analyzed single nucleotide polymorphism (SNP) data for identifying genetic susceptibility to diabetic retinopathy
• Implemented Lasso Regression and Random Forest algorithms for feature selection to identify the most relevant SNPs
• Trained and evaluated kNN, SVM, and Gradient Boosted Decision Trees for predicting disease susceptibility
• Performed data preprocessing and feature engineering on high-dimensional genetic datasets
• Compared model performance using accuracy, precision, and F1 score metrics to identify the best classification approach

This project introduced me to the intersection of machine learning and biology, teaching me how to handle high-dimensional genomic data, apply feature selection techniques, and evaluate classification models for medical applications. It also strengthened my Python coding practices and data analysis skills.

Recognition of Devnagri Script using Virtual Pen Hover Aug 2021 - Dec 2021

BITS Pilani

This project was undertaken as part of faculty-supervised research at BITS Pilani, aimed at creating a virtual hover pen system with recognition support for Devnagri script.

• Designed a virtual hover pen using OpenCV contour detection to track pen movement and capture handwritten strokes in real-time
• Built an Encoder-Decoder model for recognizing Devnagri (Hindi) script characters written with the virtual pen
• Implemented image preprocessing pipelines for cleaning and normalizing handwritten character inputs
• Integrated Hindi language recognition support into the hover pen system for end-to-end text capture
• Developed using OpenCV, Google Colab, and Jupyter Notebooks with Python-based deep learning frameworks

This project introduced me to the field of Human-Computer Interaction (HCI) and computer vision. It gave me hands-on experience with real-time video processing using OpenCV, sequence-to-sequence deep learning models, and the challenges of building interactive systems that bridge the gap between physical gestures and digital text.

Crop Disease Identification Aug 2020 - May 2021

BITS Pilani

This project was undertaken as part of research at BITS Pilani, focused on developing a novel deep learning architecture for detecting diseases in tomato leaves. The resulting work was published in Springer's Advances in Information and Communication Technology.

• Developed a novel architecture combining ResNet-50 with Inception modules and deformable convolutions for robust disease identification
• Achieved 98.16% accuracy on the PlantVillage dataset, outperforming the traditional ResNet model (97.5%)
• Improved accuracy by over 30% on real-world PlantDoc images captured in less-than-ideal field conditions
• Created a new augmented dataset by merging lab-curated and real-world farmland images to improve model robustness
• Leveraged skip connections, multi-scale filters, and deformable kernels to enhance disease recognition in realistic agricultural settings

This project taught me about the practical applications of deep learning in the agriculture sector. It strengthened my skills in designing custom CNN architectures, working with image datasets, data augmentation techniques, and evaluating model performance across different real-world conditions. It also gave me valuable experience in academic research and publishing.

Earthquake Forecasting Aug 2020 - December 2020

BITS Pilani

This project was undertaken as part of research at BITS Pilani, applying neural network models to earthquake forecasting in seismically active regions. The resulting work was published in Springer Nature's Disaster Management in Complex Himalayan Terrains.

• Built a time series forecasting model using artificial neural networks to predict earthquakes across five seismic regions including the Indian Himalayan Region
• Extracted eight key seismicity indicators based on the Gutenberg-Richter law and other empirical relations from seismic data spanning 1980 to 2020
• Achieved an accuracy of 90.4% and an F1-score of 0.89 for predicting earthquakes of magnitude higher than the threshold within 30 days
• Optimized the neural network architecture using deep learning techniques to model nonlinear and heterogeneous seismic processes
• Evaluated model performance across multiple seismic zones including the Himalayas, Central Java, Sumatra, Sulawesi, and Southeast Asia

This project exposed me to the field of geophysics and the important work being done in forecasting natural disasters. It taught me how to work with time-series data, extract domain-specific features, and apply machine learning to real-world scientific challenges. It also gave me valuable experience in academic research and publishing.

Facial Recognition Based Attendance System May 2020 - July 2020

Tamil Nadu Health Systems Project

This project was undertaken as an internship at the Tamil Nadu Health Systems Project, a government organization under the Tamil Nadu state government, aimed at reducing COVID-19 transmission risks in hospitals.

• Developed a facial-recognition based attendance system using Computer Vision to eliminate contact with shared surfaces during the pandemic
• Implemented the Haar Cascade algorithm with a four-stage methodology: Haar Feature Selection, Integral Image Creation, Adaboost Training, and Cascading Classifiers
• Built real-time facial matching against a patient database for automated registration and token number assignment
• Achieved over 95% accuracy in facial recognition and significantly reduced queue sizes and patient wait times in large hospitals
• Designed a patient calling system based on token queues to streamline the hospital registration workflow

This project taught me the practical application of computer vision using the OpenCV library and multiple approaches to the facial recognition challenge. Beyond technical skills, it strengthened my ability to work in a team, communicate ideas effectively in a professional setting, and present solutions to government stakeholders.

Epidemiological Analysis of COVID-19 March 2020 - June 2020

BITS Pilani

This project was developed as part of the Applied Stochastic Processes course at BITS Pilani, applying mathematical epidemic models to analyze COVID-19 dynamics in India.

• Modeled COVID-19 spread using the SIR (Susceptible-Infected-Recovered) epidemic model with differential equations to predict infection dynamics over time
• Implemented Discrete Time Markov Chain (DTMC) and Continuous Time Markov Chain (CTMC) stochastic models for probabilistic disease spread analysis
• Applied the Gillespie algorithm for stochastic simulation and EM method for parameter optimization
• Estimated model parameters by minimizing squared error loss and calculated the reproductive number to be approximately 1.2
• Validated model assumptions against real pandemic data from India to assess transmission patterns and recovery rates

This project deepened my understanding of stochastic processes, mathematical modeling, and their application to real-world epidemiological challenges. It strengthened my skills in MATLAB programming, differential equation solving, and parameter estimation techniques for complex dynamical systems.

English to Hindi Language Transliteration June 2020 - July 2020

BITS Pilani

This was a self-motivated deep learning project at BITS Pilani, exploring neural network architectures for character-level script conversion from English to Hindi.

• Built an Encoder-Decoder model for transliterating English text into Hindi Devnagri script at the character level
• Implemented Gated Recurrent Units (GRUs) as the core recurrent architecture for both encoder and decoder components
• Integrated an attention mechanism to enhance transliteration quality by allowing the decoder to focus on relevant input characters
• Designed data preprocessing pipelines for handling character-level mappings between English and Hindi scripts
• Evaluated model performance on sequence-to-sequence transliteration accuracy across varied input lengths and character combinations

This project strengthened my understanding of sequence-to-sequence deep learning architectures, attention mechanisms, and recurrent neural networks. It also gave me practical experience in natural language processing and the challenges of working with multilingual text data.

Occlusion Analysis & Filter Visualization March 2020 - April 2020

BITS Pilani

This was a self-motivated project at BITS Pilani, exploring CNN interpretability techniques to understand how deep neural networks make classification decisions.

• Implemented occlusion sensitivity analysis to systematically occlude image regions and measure their impact on CNN classification confidence
• Generated heat maps highlighting the most important image regions that influence the network's predictions
• Performed filter visualization to visualize what convolutional filters learn at different layers of the network
• Analyzed feature learning patterns across multiple CNN layers to understand hierarchical feature extraction
• Applied these techniques to various image classification tasks to demonstrate CNN interpretability across different domains

This project introduced me to the field of explainable AI and CNN interpretability. It taught me how to look beyond model accuracy and understand the reasoning behind neural network predictions, a skill essential for building trustworthy AI systems. It also strengthened my understanding of convolutional neural network internals and visualization techniques.