Devan Srinivasan
Hi! I'm Devan, an MSc student at the University of Toronto part of the computational linguistics lab, supervised by Dr. Gerald Penn. My research interests revolve around semantics, reasoning, and a little philosophy as well as machine learning interpetability. Irrespective of academia, I generally love building and solving meaningful problems. In my free time I like playing soccer and boxing.
CONNECT WITH ME
Some Projects
AND CONTRIBUTIONS
Reading Maturity
Description
We investigate the capability of LLMs for the novel task of literary quality assessment. We introduce a new dataset with stories collected from the Wattpad platform. We conduct experiments on both our Wattpad dataset and the Project Gutenberg dataset, comparing multiple approaches. We test a plethora of LLMs, as well as existing sentiment-based models, and introduce the application of a recurrence over the lightweight Longformer for this task. We find long context deteriorates the capability of LLMs, and that while finetuning flagship LLMs performs best, lightweight models remain competitive.
Calc4Lean
Description
We used a novel data bootstrapping technique to create a dataset of calculus problems and their solutions in Lean. We then benchmarked existing flagship LLMs on this dataset demonstrating the inability of current LLMs to solve simple problems in formal language. The repo has been deprecated after the previous submission as we are reworking the project.
Description
I sought to investigate the interpretability of diffusion models for image generation. I used a simple model, and trained it to generate digits. I then used various ML interpretability techniques to investigate any patterns and "interpretable" observations in the model. This was an undergraduate research project.
Description
Our team investigated incorporating value based gradients, and multistep prediction in the dreaming process within the DreamerV2 model. DreamerV2 was previously SOTA (replaced by DreamerV3) on long-term RL planning tasks (like Minecraft). My main contribution was the value based gradients, training and evaluating the model, and helping with the paper. This was an undergraduate research project.
Harlowe Transpiler
RepoDescription
I built a transpiler from Harlowe 3.3.6 to Javascript that generates the according JSON 6 object, using abstract syntax trees, grammar, and the ANTLR parser. It was designed to allow for cross-story Twine compilation and runtime environments. I included in the project, an example of integration with JS into an existing Twine game.
Experience
Summer 2025
Machine Learning Researcher
Internship | Pinterest
- Improving upon SOTA models by 79% and pioneering
- Shipping models to 32M+ users
- Pioneering infrastructure for VLM Fine-Tuning company-wide
2024 - 2026
MSc in NLP
University of Toronto
- Pursuing a MSc in NLP under the supervision of Dr. Gerald Penn.
- Research Interests: Computational Linguistics, Interpretability, Semantics & Philosophy
Summer 2024
Machine Learning Engineer
Enginable
- Built an enterprise level agentic automation system for our largest client, cutting 30+ hours of work to mere minutes
- Lead client conversations translating business requirements to engineering plans, so I could deliver real impact
Spring 2024
Systems Researcher
Internship | embARC
- I used GPU profilers (Nsight Systems) to analyze the memory transfer during RL training inside Nvidia's Isaac Gym
- I also learned about task planning and SLAM systems for automated chemistry labs
Fall 2023
Operations and Development
Koala Studios
- Lead our sales and marketing strategy, closed multiple clients in 2 months
- Built the blog infrastructure
Summer 2023
Software Engineer
Internship | Steamlabs
- Designed and built a new navigation & tracking system feature for the main project used by hundreds
- My new feature introduced new cross-story dynamic functionality into Twine-based games
2022 - 2023
Founder
Mathamoo
- I developed a new program for mathematics centred around learning how to learn
- All students reported notable academic improvement.
Summer 2022
Solutions Associate
Deloitte
- Founded the CDC Learning Community of over 90 members across Deloitte Montreal
- Trained and certified myself in Azure and Databricks ahead of schedule by 5 days and taught others as well
Summer 2021
Data Scientist
Internship | University of Toronto - Social Science Research Team
- Used AI APIs to review massive datasets, cutting time from 60+ hours to mere minutes
- Saved the team 3+ hours by mass-producing datasets through script automation and creating custom plots for analysis
Summer 2021
Software Engineer
Internship | University of Toronto - CS Dept.
- Improved quality of the code by utilizing react-chartjs2 library and reducing code needed
- Solved bugs to improve user experience, utilized by thousands of users
2020 - 2024
Bachelor's Degree
University of Toronto
- HBSc Computer Science Specialist & Mathematics Major
- 4.0 GPA