Dishant Sutariya
Bremen, Germany
Machine Learning Engineer and AI Researcher with 3+ years of hands-on experience building, evaluating, and deploying ML/AI systems. Master's graduate in Medical Engineering (FAU Erlangen) with a research track record including MICCAI publications and industry collaborations at Fraunhofer MEVIS. Passionate about responsible AI — combining engineering rigour with statistical validity to deliver production-ready solutions. Fluent in English, conversational German (B1).
- Designed CheXmask lung-cropping pipeline reducing racial encoding bias (AUROC 0.623→0.593) while preserving diagnostic accuracy (0.738 AUROC).
- Lead author of BVM 2026 paper on preprocessing methods for racial encoding in chest X-ray diagnosis.
- Ensured compliance with diverse patient demographics through rigorous statistical validation.
- Built Python statistical toolkit (meval) for ML model fairness evaluation — AUROC, PRG, uncertainty estimation, multiple-comparison corrections.
- Developed interactive Plotly dashboards saving 8 hours/week of manual bias review.
- Coordinated MICCAI 2025 submission; versioned experiments in DVC for a reproducible paper package on first submission.
- Resolved class imbalance in a 50,000+ record vehicle insurance dataset using SMOTE — F1+10%, false negatives −8%.
- Comparative analysis (Logistic Regression, Random Forest, XGBoost) — XGBoost selected at 92% accuracy.
- Boosted overall accuracy by 6% through systematic hyperparameter tuning and 10-fold cross-validation.
- Specialisation: Medical Image and Data Processing
- Thesis on fairness evaluation in deep learning models for medical imaging
- Foundation in algorithms, software engineering, and data structures
- meval: A Statistical Toolbox for Fine-Grained Model Performance Analysis
- Bias audits with uncertainty estimation and multiple-comparison controls
- 2-day hackathon: image registration of chest X-rays to a standard lung for lesion monitoring
- Won through democratic voting among competing teams
Open-source Python toolkit for comprehensive and statistically rigorous ML model evaluation. Implements AUROC, Precision-Recall Gain (PRG), uncertainty estimation, and multiple-comparison corrections — enabling transparent bias audits across patient subgroups. Interactive Plotly dashboards reduce manual analysis by ~8 hours/week.
Investigates how standard preprocessing steps (lung cropping, histogram normalisation) modulate racial information encoded in chest X-ray images. The CheXmask-based cropping pipeline reduces a classifier's ability to predict patient race (AUROC 0.623→0.593) while preserving diagnostic accuracy (0.738 AUROC), establishing new protocols for equitable CXR diagnosis.
Open-source Python library for fine-grained, statistically rigorous model performance analysis. AUROC, PRG, uncertainty estimation, multiple-comparison corrections, and interactive Plotly dashboards. Published at MICCAI 2025.
RAG application allowing users to upload PDF, TXT, CSV, Excel, Word, and JSON files. Creates searchable embeddings via FAISS vector store and answers questions using Groq's LLM based on the most relevant document context.
ViT-conditioned diffusion model in PyTorch to synthesise multi-energy PCCT images from 50 keV input. Multi-objective loss (MSE + SSIM + perceptual) achieves Dice score 0.86, preserving structural detail and material discrimination.
Deep learning system that mitigates racial demographic bias in chest X-ray diagnostic models. Implements preprocessing pipelines and bias metrics to produce equitable predictions across patient demographics.
End-to-end ML pipeline for wine quality prediction — data preprocessing, model training, and inference service. Containerised with Docker and hosted on AWS EC2. Demonstrates full MLOps lifecycle from experiment to production.