AI safety researcher investigating how large language models internally represent safety-critical concepts. I want to understand whether alignment techniques genuinely shape model internals or simply produce surface-level behavior, and how mechanistic interpretability can help us build more robust safeguards.
Completed the Bluedot Impact Technical AI Safety Course (2026) and the project sprint, replicating and extending refusal directionality findings in Llama-3.1-8B-Instruct. Read the write-up.
Experience
AI/ML Research Intern
Vizuara AI Labs · India · Jan 2026 – Mar 2026
- Built and evaluated a citation-grounded RAG system for OpenFOAM technical documentation using hierarchical chunking, dense retrieval, and cross-encoder reranking; achieved 4.6/5 overall generation quality and 5.0/5 citation correctness. Accepted at ACL 2026 RAG4Report Workshop.
- Fine-tuned RF-DETR for multi-class SAR object detection on a custom dataset, achieving 0.614 mAP and improving detection robustness across heterogeneous scene categories.
- Developed an agentic VLM pipeline that transforms research PDFs into structured technical digests by generating narrative summaries, AI-generated diagrams, and executable Jupyter notebooks through automated multimodal reasoning workflows.
Aerodynamics Research Assistant
San Diego State University · USA · Aug 2021 – Dec 2023
- Ran 3D CFD simulations on an HPC cluster (SLURM), cutting wall-time by 75% through parallelization; processed and visualized large-scale flow-field datasets using Python and ANSYS CFD-Post.
- Quantified surface roughness effects through wind tunnel experiments and CFD: measured a 5° delay in flow separation and a 16% lift increase, contributing to a co-authored AIAA 2025 publication.
Skills
Core ML/AI
Mechanistic Interpretability, LLM Evaluation & Benchmarking, Post-Training Alignment, Machine Learning, Deep Learning, Computer Vision, NLP, Reinforcement Learning, RAG, Object Detection, Image Segmentation
Languages
Python, C++, Bash/Shell, MATLAB
Frameworks & Libraries
PyTorch, TensorFlow, LangChain, FastAPI, OpenCV, Hugging Face, Unsloth, ChromaDB
MLOps & Infrastructure
MLflow, Databricks, Docker, Git, Weights & Biases, HPC/SLURM
Projects
Geometry of Refusal: Mechanistic Interpretability of Safety-Relevant Representations
Replicated and extended Arditi et al. (2024) on Llama-3.1-8B-Instruct. Extracted per-layer DIM refusal directions across 32 layers; layer 12 was the strongest. Both single-direction and per-layer ablation achieved 100% flip rate on training prompts (0/150 random baseline). On JailbreakBench (n=100), true ASR was 100% with coherence intact across all conditions. Applying layer 12's direction across all layers dropped refusal score to -3.67; skipping it matched per-layer (-0.96), confirming refusal is distributed. Activation addition on harmless prompts pushed refusal score 9x (1.48 → 13.46). Bluedot Impact AI Safety project sprint.
Blog
Customer Churn: End-to-End ML Pipeline
Production ML pipeline for customer churn prediction with model versioning, automated retraining, and inference API using MLflow, FastAPI, and Docker.
Repo
OpenFOAM RAG: Citation-Grounded Q&A
Citation-grounded RAG system for OpenFOAM technical documentation using hierarchical chunking, dense retrieval, and cross-encoder reranking. Achieved 4.6/5 generation quality and 5.0/5 citation correctness. Published at RAG4Report Workshop @ ACL 2026.
Repo · Live
PaperDigest: Agentic PDF-to-Digest
Agentic system that ingests academic PDFs, extracts structured content, and generates concise digests using FastAPI and Gemini. Automated pipeline from upload to summary with key-finding extraction.
Profile
SAR Aircraft Detection (RF-DETR)
Fine-tuned RF-DETR on synthetic aperture radar imagery for aircraft detection. Achieved mAP 0.614, competitive with published benchmarks. Diagnosed overfitting and scoped SARDet-100K for data scaling.
Profile
Student Misconception Detection
Fine-tuned Qwen2.5-Math-1.5B-Instruct with Unsloth and a compact MLP classification head to identify student misconceptions in math reasoning. Improved MAP leaderboard from 0.885 to 0.925 (+4.5%).
Repo
Stable Video OCR: License Plate Recognition
Real-time detection using fine-tuned YOLO11 + EasyOCR. Eliminated flickering via a majority-voting algorithm over a 12-frame rolling buffer, producing stable text output across video frames.
Repo
Video Instance Segmentation: Mask R-CNN
Per-frame object detection and pixel-level segmentation pipeline using PyTorch's pre-trained Mask R-CNN. Renders per-instance colored masks and bounding boxes on live video streams.
Repo
Research
Decompose, Retrieve, Cite: A RAG Pipeline for Structured Report Generation from Technical Documentation
RAG4Report Workshop @ ACL 2026 — Accepted
Himanshu C. Dhurve, et al.
Effect of Surface Texture on the Lift and Drag of Small Spinning Balls
AIAA Aviation Forum & Ascend 2025 — Published
Joseph Katz, Himanshu C. Dhurve
Read Paper
Resume
Education
M.S. Aerospace Engineering
San Diego State University, USA · GPA 3.52 / 4.0 · 2021 – 2024
Recipient, Master's Research Scholarship · 2022
B.Tech Mechanical Engineering
Dr. Babasaheb Ambedkar Technological University, India · GPA 8.57 / 10.0 · 2014 – 2018