Skip to content
View skn8733's full-sized avatar
1 follower · 4 following

Block or report skn8733

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
skn8733/README.md

Shasanka Nudurupati

MS Cybersecurity @ NYU Tandon (3.8 GPA) · BS Math & CS @ UC San Diego

I build ML systems with a focus on reliable pipelines — the data processing, tokenization, and evaluation work that determines whether a model actually learns the right thing. Most of my debugging happens at the tensor level, not the architecture level.

Projects

Pixels to Predictions — Fine-tuned SmolVLM-500M using QLoRA for multimodal science QA. Found and fixed silent loss masking and tokenizer alignment bugs that wasted days of GPU compute. 10th/54 on Kaggle (0.822).

Defection Probe Generalization · LessWrong writeup — Tested whether Anthropic's sleeper agent detection technique generalizes to naturally-arising deception. Ran Llama 70B inference across 90 prompt-conditions on rented A100s, manually scored every output. Honest null result. BlueDot Impact AI Safety Sprint.

Offline RAG Agent — End-to-end retrieval pipeline over expert call transcripts using local Mistral 7B, ChromaDB, and LangChain. No cloud APIs. Identified and documented hallucination patterns from cross-transcript context contamination.

Python Vulnerability Detection — ML pipeline classifying 14 insecure Python code patterns using AST-based data augmentation, TF-IDF vectorization, and Logistic Regression/XGBoost.

Pinned Loading

  1. transcript-answer-RAG-agent transcript-answer-RAG-agent Public

    End-to-end retrieval pipeline over expert call transcripts using local Mistral 7B, ChromaDB, and LangChain. No cloud APIs. Identified and documented hallucination patterns from cross-transcript con…

    Python

  2. defection-probe-generalization defection-probe-generalization Public

    Tested whether Anthropic's sleeper agent detection technique generalizes to naturally-arising deception. Ran Llama 70B inference across 90 prompt-conditions on rented A100s, manually scored every o…

    Jupyter Notebook

  3. QLoRA-Fine-Tuning-of-SmolVLM QLoRA-Fine-Tuning-of-SmolVLM Public

    VLM for scientific multiple-choice reasoning, combining image understanding with textual context (question, lecture, hint, solution) to predict the correct answer index for each example.

    Jupyter Notebook