Kaushik Chatterjee Email & Phone Number

Data and Privacy Protection Lab 🤖- Researching data-efficient alternatives to fine-tuning LLMs for code transpilation- Augmenting LLM-as-a-Judge methods with constrained decoding for domain-specific tasks- Developing faster, more adaptable RAG systems via knowledge graphs

• Facilitated the E2E data science lifecycle for time-series sensor data, performing EDA, model selection, and system deployment.• Trained autoregressive linear and decision-tree-based models in ML.NET for real-time sensor data, as well as performed feature engineering and cross-validation to improve performance.• Assessed Open-source LLMs with varying quantizations for RAG benchmarking in a healthcare chatbot.• Red-teamed LLMs to identify vulnerabilities & biases and implemented appropriate guardrails using NVIDIA's NeMo library.

• Presented a real-time stock-trading bot at CUCAI, Canada's largest undergraduate conference for AI. • Developed an ANN in PyTorch to perform sentiment analysis on news data for generating trading signals using pre-trained GloVe embeddings, achieving a validation accuracy of 0.879.• Leveraged batch normalization layers and dropout to accelerate training time while mitigating overfitting, increasing model performance by 89%.

Delivered a presentation to a room of ~30 students, ranging from undergraduates to post-graduate researchers, on AI ethics and model interpretability, covering the timeline of AI models, interpretability techniques for LLMs, and core concerns and considerations of AI alignment. Additionally, I created a Colab to allow participants to directly apply these techniques and understand their importance.

• Leveraged NLP methods and large language models to tackle real-world challenges for clients, including facilitating the development of a retrieval augmented generation model for a multinational telecommunications company resulting in over $1.8m in savings.• Productionized a bespoke language model by deploying it to a REST API, helping to create an end-to-end MLOps pipeline.• Developed a cross-validation pipeline for retrieval models in addition to a package for quantitatively evaluating language models for domain-specific tasks, soon becoming known as "The Evaluation Man" :)

Hack the Woodz is a 24-hour, beginner-focused virtual hackathon & ideathon that featured a Pitch-Only track, a Coding track, and a Hardware track, ensuring anyone with a passion for technology or creation was able to participate.As Co-Founder, I led discussions with key stakeholders in order to organize the event, including reaching out to sponsors and negotiating the logistics with our fellow organizers. I marketed the event through various social media platforms, using targeted campaigns in order to reach a broad audience and achieve our goals.During the hackathon, I was in charge of organizing various events and providing participants with assistance and guidance. I led interactive GitHub and Artificial Intelligence workshops where I taught participants the core concepts in an approachable, dynamic manner while engaging them in discussions regarding the topic material. I also was in charge of overseeing the logistics of the event and ensuring that the submitted projects were compliant with our rules. Finally, I conversed with our judges in order to organize the rankings and handled aspects of the prize distribution.As a result of the work of my team and I, Hack the Woodz garnered 144 participants and over $33,000 CAD in sponsorships.

Bachelor Of Computer Science, Cumulative Gpa: 3.85/4.00

Activities and Societies: Winter'24: Machine Learning Engineer, WAT.ai Fall'23: Workshop Lead, Data Science Club; Volunteer Lead, Hack.

Certificate, Quantum Computing, 11

Qubit by Qubit's Introduction to Quantum Computing is the first-ever year-long quantum course for 5,000 high school students and above. It.