Building trustworthy AI and privacy-preserving systems for healthcare, security and real-world applications.
I’m Suryansh Maurya, a Computer Science and Engineering undergraduate graduating in 2028.
My experience lies at the intersection of artificial intelligence, privacy, healthcare technology, backend engineering and trustworthy machine learning.
- 🧬 Google Summer of Code 2026 Contributor at Healthy.inc/Alaska
- 🔐 Secure Machine Learning Research Intern at IIT (BHU)
- 📄 Research Collaborator with NTU Singapore and the University of Warwick
- 📰 Published researcher in Transactions on Machine Learning Research
- 🩻 Open-source contributor to privacy-sensitive medical-imaging systems
- ⚙️ Full-stack developer and AI/ML engineer
- 🏆 Winner of the Wells Fargo Ideathon
| 350+ hours GSoC engineering |
TMLR 2026 Published research |
23% faster TabFlowM training |
871 subjects ABIDE-I research |
My latest experience, education, research, projects and technical skills are available in my resume.
Google Summer of Code Contributor | May–September 2026
Selected for Google Summer of Code 2026 to contribute to Bio-Block, a privacy-preserving healthcare-data platform.
My work focused on building secure, privacy-aware and production-oriented infrastructure for processing sensitive medical information.
- Invested more than 350 hours in project development and research.
- Developed privacy-preserving medical-data ingestion workflows.
- Implemented PHI anonymization using Microsoft Presidio.
- Built OCR-based redaction pipelines for sensitive information embedded in medical documents.
- Worked with Mondrian k-anonymity to strengthen privacy protection.
- Developed backend services and APIs using Python and FastAPI.
- Built hybrid medical-information retrieval workflows.
- Combined BM25, sentence embeddings, ChromaDB, Reciprocal Rank Fusion and Maximum Marginal Relevance.
- Contributed to medical-image preview workflows for WSI and NIfTI files.
- Worked with IPFS and blockchain-backed healthcare-data workflows.
- Improved CI/CD stability and resolved integration issues.
Python FastAPI Presidio OCR ChromaDB BM25 RRF MMR IPFS Docker GitHub Actions
November 2025–July 2026
Worked on secure and trustworthy machine-learning methods for privacy-sensitive data.
My research examined how machine-learning systems can preserve privacy without sacrificing predictive performance, data utility or fairness.
- Synthetic-data generation
- Federated learning
- Differential privacy
- Trustworthy artificial intelligence
- Privacy-preserving machine learning
- Fairness-aware evaluation
- Secure data sharing
- Evaluated synthetic datasets using classification performance, privacy and utility.
- Compared models trained on real and synthetic data.
- Investigated privacy risks introduced by data-generation techniques.
- Evaluated model fairness across different groups.
- Conducted ablation studies to measure individual system components.
- Studied trade-offs between predictive accuracy, privacy protection and statistical similarity.
- Analysed minority-group representation and model fairness.
- Worked on research for privacy-sensitive domains such as healthcare and finance.
Python PyTorch TensorFlow Scikit-learn Pandas NumPy Differential Privacy Federated Learning Synthetic Data
January–August 2026
Collaborated on research involving generative modelling for mixed-type tabular data.
This collaboration resulted in TabFlowM, published in Transactions on Machine Learning Research.
- Co-developed a lightweight flow-matching method for tabular-data synthesis.
- Worked with continuous and categorical variables in a shared latent representation.
- Studied latent-space generative modelling for heterogeneous datasets.
- Participated in experimentation, evaluation and research writing.
- Compared TabFlowM with established tabular generative models.
- Achieved the best average rank in the reported experimental evaluation.
- Demonstrated approximately 23% faster training than TabDiff.
- Evaluated statistical quality and downstream machine-learning utility.
Flow Matching Tabular Synthesis Latent Representations Generative AI Mixed-Type Data Trustworthy ML
Contributed to a privacy-sensitive healthcare system connecting clinical imaging infrastructure with downstream deep-learning workflows.
- Implemented real-time DICOM retrieval.
- Worked on medical-image metadata extraction.
- Developed anonymization workflows for sensitive clinical information.
- Integrated medical-imaging workflows with Orthanc PACS.
- Worked with medical-imaging datasets from The Cancer Imaging Archive.
- Supported downstream deep-learning and medical-image analysis.
- Contributed fixes across backend logic, data processing and system integration.
- Worked across multiple healthcare-data and medical-imaging workflow areas.
MATLAB DICOM Orthanc PACS TCIA Medical Imaging Deep Learning Healthcare Privacy
Worked on machine-learning research involving neuroimaging, behavioural signals and healthcare analytics.
For an ABIDE-I autism study, I worked with data from 871 subjects:
- 418 subjects with ASD
- 453 control subjects
- Achieved 0.74 ± 0.04 AUC
- Achieved 0.70 ± 0.04 accuracy
- Identified three neurobiological subtypes
The research involved evaluating machine-learning performance while accounting for variability, uncertainty and clinically meaningful patterns.
Healthcare ML Neuroimaging Autism Research Clustering Classification Statistical Evaluation
TabFlowM: Lightweight Flow Matching for Mixed-Type Tabular Data Synthesis in Latent Space
Transactions on Machine Learning Research, July 2026
TabFlowM is a lightweight generative framework for synthesizing mixed-type tabular data in latent space.
- Supports heterogeneous continuous and categorical features.
- Uses flow matching for efficient generative modelling.
- Achieved the best average rank across the reported benchmarks.
- Trained approximately 23% faster than TabDiff.
- Evaluated using statistical similarity and downstream machine-learning utility.
- Designed to improve the efficiency of high-quality tabular-data synthesis.
Flow Matching Tabular Data Generative Modelling Latent Space Synthetic Data
🛡️ RiskLoomAI-native payment-risk intelligence system RiskLoom is a secure Razorpay-oriented fraud-risk platform featuring:
|
Privacy-preserving healthcare infrastructure A healthcare-data platform developed through Google Summer of Code.
|
Flow Matching Generative AI Synthetic Data Federated Learning Differential Privacy Healthcare ML
- 🏆 Winner — Wells Fargo Ideathon
- 🎤 Selected Speaker — GSoC 2026 Contributor Lightning Talk
- 🔬 Published researcher in Transactions on Machine Learning Research
- 🌍 OSC Global 2026
- 💻 Elite Coders Winter of Code 2026
- 🧬 Successfully completed Google Summer of Code 2026
- 🩻 Contributor to privacy-sensitive healthcare systems



