About Me
I am a student researcher at the Johns Hopkins Center for Language and Speech Processing focusing on LLM reasoning, and a computer science student at Johns Hopkins University specializing in NLP, AI, and ML.
Publications
DeonticBench: A Benchmark for Reasoning over RulesPreprint
In: arXiv, 2026.
BibTeX
@misc{dou2026deonticbenchbenchmarkreasoningrules,
title={DeonticBench: A Benchmark for Reasoning over Rules},
author={Guangyao Dou and Luis Brena and Akhil Deo and William Jurayj and Jingyu Zhang and Nils Holzenberger and Benjamin Van Durme},
year={2026},
eprint={2604.04443},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2604.04443},
}
SocialNLI: A Dialogue-Centric Social Inference DatasetPreprint
In: arXiv, 2025.
BibTeX
@misc{deo2025socialnlidialoguecentricsocialinference,
title={SocialNLI: A Dialogue-Centric Social Inference Dataset},
author={Akhil Deo and Kate Sanders and Benjamin Van Durme},
year={2025},
eprint={2510.05458},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2510.05458},
}
QAagent: A Multiagent System for Unit Test Generation via Natural Language Pseudocode (Student Abstract)Proceedings Article
In: Proceedings of the AAAI Conference on Artificial Intelligence, 2025.
BibTeX
@article{Deo_2025,
title={QAagent: A Multiagent System for Unit Test Generation via Natural Language Pseudocode (Student Abstract)},
volume={39},
url={https://ojs.aaai.org/index.php/AAAI/article/view/35246},
DOI={10.1609/aaai.v39i28.35246},
number={28},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Deo, Akhil},
year={2025}, month={Apr.},
pages={29345-29347}
}
Feasibility of Mobile Application for Surgical Robot TeleoperationProceedings Article
In: Hamlyn Symposium on Medical Robotics, pp. 121-122, 2023.
BibTeX
@inproceedings{Deo2023,
series = {HSMR2023},
title = {Feasibility of Mobile Application for Surgical Robot Teleoperation},
url = {http://dx.doi.org/10.31256/HSMR2023.63},
DOI = {10.31256/hsmr2023.63},
booktitle = {Proceedings of The 15th Hamlyn Symposium on Medical Robotics 2023},
publisher = {The Hamlyn Centre, Imperial College London London, UK},
author = {Deo, Akhil and Kazanzides, Peter},
year = {2023},
month = jun,
collection = {HSMR2023}
}
Projects
Deontic Reasoning in Language Models (Advisor: Ben Van Durme)
Johns Hopkins Center for Language and Speech Processing
November 2025 – May 2026
- Contributed to DeonticBench, a 6,232-task benchmark evaluating how LLMs reason over obligations, permissions, and prohibitions. The benchmark spans U.S. federal tax, airline baggage, immigration, and state housing law.
- Built data-filtering pipelines to curate high-quality tasks for the benchmark's hard evaluation sets.
Social Reasoning in Language Models (Advisor: Ben Van Durme)
Johns Hopkins Center for Language and Speech Processing
September 2024 – October 2025
- Researched social natural language inference (NLI) to improve how LLMs interpret elements like sarcasm, humor, and irony.
- Created SocialNLI, the first dialogue-centric social-inference dataset, comprising 243 Friends TV transcripts, 5.3K generated inferences, and a 1.4K human-annotated eval split - grounding theory-of-mind reasoning in sarcasm and irony.
- Developed efficient counterfactual-reasoning pipeline to assign calibrated plausibility scores to social inferences, surfacing latent LLM failure modes in social reasoning, using PyTorch, vLLM, Huggingface, and Langchain.
- Constructed taxonomy of language model failure modes on dialogue-centric data.
- Leveraged token-level and rubric-based rewards to enhance reasoning processes in LLMs using reinforcement learning.
Nuss Bar (Advisor: Peter Kazanzides)
Johns Hopkins Laboratory for Computational Sensing and Robotics
February 2024 - July 2024
- Creating desktop planning and visualization app to aid clinicians in shaping bar for Nuss Procedure, used to correct Pectus Excavatum
- Conceived user study comparing efficacy of Nuss bar prototypes shown in augmented reality vs. 3D printed bars
- Created an augmented reality workflow to aid clinicians in performing Nuss procedure, potentially impacting 1000+ Pectus Excavatum surgeries annually, using Unity, C#, Python, Slicer3D, and MRTK3.
SurgiSimulate (Advisor: Peter Kazanzides)
Johns Hopkins Laboratory for Computational Sensing and Robotics
February 2022 - July 2023
- Enabled remote surgical training by turning an iPhone into a hand-held controller for the dVRK surgical robot (Swift/Obj-C and ARKit).
- Constructed a mobile application using Swift and Objective C to control a da Vinci Research Kit (dVRK)
- Leveraged ARKit to capture transformation of mobile devices and ROS and Python for robot control
- Designed and ran a user study with 16 participants to compare performance of mobile app with existing input devices, validating feasibility for mobile teleoperation and remote surgical training.
Work Experience
Software Development Engineering Intern
Amazon Web Services
May 2025 - August 2025
- Designed and built a recording system capturing user interactions in browser and automatically generates reproducible workflow, improving automation accuracy by 20% and eliminating manual workflow design, using TypeScript and Python.
Technical Advisor Intern - GenAI
Scale AI
November 2024 - February 2025
- Solved Olympiad-level competitive programming (CP) problems that only ~25% of competitive programmers can solve, to support training a code-reasoning LLM.
- Crafted specialized prompts enabling LLMs to solve Olympiad-level competitive programming problems.
Software Development Engineering Intern
Amazon AGI
May 2024 - August 2024
- Devised and developed multi-agent small language model (SLM) frameworks to enhance LLMs' advanced planning abilities using Amazon Nova and Anthropic Claude. Achieved a 220% increase over baseline on the TravelPlanner dataset.
- Implemented an evaluation pipeline for multi-agent experiments with RxJava and Python, speeding up evaluations by 400%.
Founding Engineer
Quantable.io
December 2023 - July 2024
- Constructed the site's PostgreSQL database schemas and created efficient APIs in Node.js, a custom LaTeX rendering library, admin tools, and production-ready RBAC system, to serve 1,200+ quant finance problems for 2,500+ users.
Software Engineering Intern
PayPal
May 2023 - August 2023
- Redesigned monetary transactions API with an eventually consistent data strategy, caching transactions in a local database to eliminate multiple mid-tier API calls for 100,000+ daily transactions.
- Improved API performance, reducing latency by ~100ms per API call, leveraging Java, Spring Boot, and SQL.
Course Assistant
Johns Hopkins University
August 2022 - December 2025
- Taught and evaluated 200+ students for 3 classes: AI, Data Structures and Object Oriented Software Engineering.
Achievements
- Pistritto Fellowship
- JHU Machine Learning Best Project Award
- Pava Center for Entrepreneurship Ignite Award
- JHU Whiting School of Engineering Undergraduate Conference Travel Grant
- JHU Student Sponsorship Initiative Award
- Ongoing Venture Prize and Most Creative Use of Twilio at HopHacks Fall 2021