Dhivya Sreedhar

I'm a Master's student in Information Systems with a specialization in Machine Learning and NLP at Carnegie Mellon University's Language Technologies Institute. Previously, I earned my B.E. in Computer Science (with Honors) from Anna University in 2022. My interests span the intersection of LLMs, Multimodal ML, Agentic AI, and Responsible AI, with a strong passion for taking cutting-edge research to real-world production.

At CMU, I’m collaborating on multimodal Chain-of-Thought (CoT) reasoning under Prof. Bhiksha Raj and interning at Reclamation Factory, where I’m building vision systems for plastic waste classification using NIR, XRF, and CNNs. I’m also an Applied AI Research Intern at Scale AI, evaluating LLM reasoning abilities (GPT-4, Claude, LLaMA) and fine-tuning models for QA, summarization, and code generation.

Previously, I worked at BNY Mellon on RLHF-based conversational AI optimization and knowledge distillation frameworks, achieving $4.2M in cost savings. I also spent 2 years as a full-stack Software Engineer at Zoho, building real-time anomaly detection pipelines and distributed systems for cloud security.

Outside of AI, I enjoy fostering animals, exploring nature, and lifting weights. I’m actively seeking Machine Learning and Applied Scientist roles starting January 2026 — let’s connect!

Email 1  /  Email 2  /  Resume  /  Github  /  LinkedIn  /  Scholar

CMU
MS in Information Systems, specialization in Machine Learning and NLP
Aug 2024 - December 2025

BNY
Machine Learning Engineer
January 2025 - May 2025

Scale AI
Machine Learning Intern
May 2025 - Present

Anna University
B.E. Computer Science Engineering
August 2018 - May 2022

Scale AI
Applied AI Research Intern – LLM Reasoning (May 2025 – Present)

– Led large-scale evaluation of frontier LLMs (GPT-4, Claude, LLaMA) on tasks like multi-hop QA and long-context summarization to benchmark emergent reasoning capabilities.
– Fine-tuned instruction-following models using SFT, LoRA, and DPO for math, code, and QA tasks—boosting coherence and factuality by 18.7%.

Reclamation Factory (CMU Robotics Startup)
Computer Vision and Robotics Intern (May 2025 – Present)

– Built and deployed a multi-modal plastic classification system (NIR, XRF, RGB) on NVIDIA Jetson AGX Orin with 93.5% accuracy.
– Applied ViT fine-tuning, domain adaptation, and self-supervised pretraining to cut lighting-sensitive error rates by 27%.
– Optimized inference with TensorRT, ONNX, and CUDA to achieve <50ms latency for embedded robotic use.

Bank of New York Mellon
Applied Scientist (Jan 2025 – May 2025)

– Implemented human-in-the-loop RLHF workflows for conversational AI, reducing hallucinations by 43%.
– Scaled personalized, multi-channel deployment of 9 core AI capabilities via knowledge distillation and parameter-efficient fine-tuning, saving $4.2M/year.

Zoho Corporation – ManageEngine
Applied ML Scientist (Aug 2022 – Aug 2024)

– Designed anomaly detection for cloud log analytics using z-score, Isolation Forests, EWMA, and autoencoders.
– Developed HTTP Event Collector module for low-latency ingestion across US/UK data centers.

IIITDM Kancheepuram
Research Intern (May 2020 – May 2021)
– Conducted a comparative analysis of CNN architectures (DenseNet, MobileNet, Xception, EfficientNet) for face forgery detection under the guidance of Dr. Sivaselvan B.
– Used 140k real and synthetic facial images (Flickr + StyleGAN) resized to 256×256 px for robust evaluation.
– Achieved 97% accuracy and 0.937 Matthews Correlation Coefficient with XceptionNet, improving content authentication efficiency and reducing verification time by 35%.

National Institute of Technology Calicut
Summer Research Intern (Jun 2021 – Aug 2021)
– Worked under the supervision of Dr. Prabu Mohandas on a Music Instrument Recognition system for automatic labeling of monophonic audio files.
– Converted audio signals into 2D mel-spectrograms for model input, eliminating the need for manual annotation.
– Achieved 99.17% accuracy with a CNN model and 97% with k-NN, outperforming baseline architectures through hyperparameter tuning.

Typing ReImagined: emg2qwerty
2024 – Present

Developed a real-time neuromusculoskeletal interface translating surface EMG signals into text for AR/VR platforms (Apple Vision Pro, Meta Quest). Achieved <5% CER and <30ms latency using a hybrid Conformer-Transformer with CTC loss. Integrated Flan-T5 and GPT-4 Turbo for autocorrection with SpecAugment and causal modeling for robust, low-latency inference.

AI Interview Agent with LangChain + LLMs
2024 – Present

Built a generative agentic system simulating interviews using Llama 3, LangChain, and voice interaction. Vectorizes resumes, matches job descriptions, generates dynamic questions, and evaluates responses using an LLM-as-a-judge module with contextual reasoning and scoring.

LLM-Powered Dataset Auditor
2024 – Present

Built a RAG-based label validation system for CV datasets. Used CLIP + FAISS to retrieve visual neighbors and GPT-4/Claude for contextual audit reasoning in LangGraph workflows. Logged corrections to structured formats for scalable validation pipelines.

MyTorch: Custom Deep Learning Library
2023

Implemented PyTorch from scratch using Python and NumPy. Developed autograd engine, loss functions, optimizers, layers (Linear, Conv2D, LSTM, GRU, BatchNorm), and trained models like CNNs, RNNs, GNNs, and GANs using the library. Built for educational and experimental research use.

Last updated July 2025. Thanks Jon Barron for the template!