Machine learning engineer. I train and fine-tune models, wire up the agents that use them, and build the data pipelines underneath. Just finished my M.S. at Stevens and I'm currently doing LLM-agent automation at a nonprofit.

Portfolio: anirudhksharma.com

FraudSight AI. Agentic pipeline that checks images and PDFs for AI-generated artifacts and document forgery. A vision agent does the forensic pass, then Qwen-VL-Plus, DeepSeek R1, and GLM 4.6 vote on the result. 72% accuracy on real claims; prompt routing cut token cost 50-75%.

Agent reliability (Stevens research). Worked on ShoppingBench, a 2.5M-product agent benchmark (arXiv:2508.04266). Adding strictly-typed tool schemas raised Qwen-2.5-72B's format score from 0.47 to 0.72 and removed all parameter hallucinations, at the cost of extra retries I traced in an ablation. Built a Streamlit dashboard to compare runs.

Diabetic retinopathy. EfficientNet-B0 on 143K retinal images (~22 GB), 87% balanced accuracy. A PySpark Random Forest baseline on a small cluster cut training time by more than half.

Adversarial robustness. FGSM attacks on CIFAR-10. A CNN dropped from 78.6% to about 2.3% accuracy under a small perturbation; min-max adversarial training got most of it back.

Python, C++. PyTorch and TensorFlow. PySpark, Pandas, Polars when data outgrows one box. AWS, Azure, GCP, FastAPI, Postgres, Elasticsearch. MCP for agent tooling lately.

Portfolio · LinkedIn · Medium · anirudhksharma02@gmail.com