Akul Nehra
Built a federated ensemble learning framework that jointly optimizes privacy, fairness, and calibration for financial risk assessment across 24 federated clients. Achieved 91.8% accuracy on 150,000 real-world samples while satisfying GDPR, ECOA, and Basel III compliance. Published at IEEE RCSM 2025.
Where Akul Started
His Background
- • 9th grader at Genesis Global School, Noida (IB Curriculum)
- • Accepted to Harvard HUVTSP and Stanford AI4ALL
- • Built KrishiAI, an IoT model providing AI-driven insights for farmers
- • Proficient in Python, scikit-learn, Pandas, Flask, SQL
- • Top speaker and winning team at Malaysia International WSDC debate
- • Recommended for IB Youth Action Fund
His Goals
- • Publish a peer-reviewed research paper
- • Apply AI and ML to real-world financial problems
- • Build a competitive profile for top university admissions
- • Present at science fairs and conferences like ISEF
- • Create research with positive community impact
The Problem He Wanted to Solve
Financial institutions hold vast credit risk datasets but regulatory frameworks like GDPR prohibit data centralization. Existing federated learning approaches address privacy, fairness, or calibration in isolation, but no framework simultaneously optimizes all three for heterogeneous ensemble architectures. Akul wanted to build the first system that does, making AI-driven financial risk assessment both accurate and regulatory-compliant.
The Research
Akul developed a federated ensemble learning framework combining five heterogeneous base learners (logistic regression, random forest, gradient boosting, SVM, and multilayer perceptron) across 24 federated clients. The framework uses differential privacy with Renyi DP accounting, meta-learning aggregation via stacking, and secure proxy training for non-aggregable models.
Privacy-Preserving Federated Ensemble Learning for Fair and Calibrated Financial Risk Assessment
No existing framework simultaneously addresses privacy, fairness, and calibration for heterogeneous federated ensembles in financial applications
5 heterogeneous models across 24 clients with differential privacy, meta-learning stacking, and secure proxy training
150,000 real-world GMSC loan samples + 45,000 synthetic samples across 24 federated clients with non-IID heterogeneity
91.8% accuracy (1.7pp gap from centralized), ECOA fairness (DP gap < 0.05), Basel III calibration (ECE < 0.05), formal differential privacy
Regulatory-Compliant AI from a 9th Grader
What makes Akul's work exceptional is its real-world regulatory relevance. His framework simultaneously satisfies GDPR Article 32 (formal differential privacy), ECOA anti-discrimination requirements (demographic parity gaps below 0.026), and Basel III calibration standards (ECE below 0.032). He also discovered that differential privacy acts as regularization, improving accuracy by +0.04pp on synthetic data, challenging conventional privacy-utility trade-off assumptions. The results were confirmed with rigorous statistical testing (t = 32.516, p = 0.0009).
Accuracy (GMSC)
Real-World Samples
Federated Clients
Compliance Standards
The Outcome
Full Paper Accepted at IEEE Conference
IEEE RCSM 2025
Full Paper
Privacy-Preserving Federated Ensemble Learning for Fair and Calibrated Financial Risk Assessment
Accepted for Publication
Strong Python/ML skills with experience building AI models, but no peer-reviewed publications or formal research credentials
First-author IEEE publication on privacy-preserving federated learning, with a framework that satisfies three major regulatory standards simultaneously
The Bigger Picture
Years old when he built a regulatory-compliant federated AI framework and published at an IEEE conference
Gap from centralized performance while maintaining formal differential privacy across 24 federated institutions
Framework to simultaneously optimize privacy, fairness, and calibration for heterogeneous federated ensembles in finance
Ready to Start Your Research Journey?
Join the YRI Fellowship and work with expert mentors to conduct original research, get published, and build a profile that stands out.
Apply Now