AI risk assessment models often bring preconceived biases from datasets, resulting in unfair pricing or claims refusals.
Design and implement a fairness aware AI model for dynamic insurance pricing.
Use bias detection metrics, and apply mitigation measures (i.e., pre-processing, in-processing, post-processing).
Assess effectiveness of model, under trade-offs between fairness and accuracy.
Use synthetic datasets, or open insurance datasets.
Implement using Python (Scikit-learn, AIF360).
Quantitatively assess biased vs. de-biased models using fairness metrics (Demographic Parity, Equal Opportunity).
Fairness in Machine Learning, Algorithmic Bias (Barocas et al., 2016).
Real-time insurance fraud detection with Edge AI and Federated Learning
To design an AI model that is compatible with edge-based systems for the intended purpose of detecting fraud in vehicle or health insurance.
Investigate using federated learning which will allow organizations to learn on their own data without the direct sharing of raw data.
Building an edge-based environment (such as mobile telematics).
Use frameworks like TensorFlow Federated and PySyft.
Assess the model’s accuracy, latency and communications constraints.
Federated Learning (Google, 2017), Edge AI, Adversarial ML risk evaluation.
K. Kuppan et al., 2024.
Q. Yang, Y. Liu, T. Chen and Y. Tong, “Federated Machine Learning: Concept and Applications,” ACM Transactions on Intelligent Systems and Technology, vol. 10, no. 2, pp. 1–19, 2019.
Y. Kang, Y. Li and Z. Xu, “Edge Computing for Real-Time AI Applications,” IEEE Internet of Things Journal, vol. 7, no. 2, pp. 1564–1574, Feb. 2020.
Contract review of real estate transactions is often a manual and error-prone process, yet compliance auditing using AI remains scarcely explored.
Build an NLP model to identify compliance risks in property leases or contracts.
Automate clause detection and red-flagging using rules for regulations (e.g., tenancy, zoning).
Fine-tune pre-trained models (BERT, LegalBERT) on labeled legal documents datasets.
Apply rule-based post-processing for tagging by contract-specific.
Tools: Python, Hugging Face Transformers, SpaCy.
Legal NLP, Automation of Compliance, Document Intelligence.
K. Kuppan et al., 2024.
D. Chalkidis et al., “Legal-BERT: The Muppets straight out of law school,” in Proc. Findings of EMNLP, 2020.
M. Zhong, H. Chen, C. Yu and H. Xie, “Smart Contract Semantic Analysis Based on NLP,” in IEEE Access, vol. 8, pp. 171448–171459, 2020.
Current models for pricing real estate rentals ignore multimodal signals (e.g., text, images and geographic data) when pricing real estate dynamically.
Develop a multimodal AI model that uses images of real estate rental properties, text descriptions and location to set a rental price at a point in time.
Extend this pricing signal to account for demand signals in the market that will vary continuously over time.
Generate CNN outputs as image features; NLP outputs as features for the property description; integrate GIS data on location.
Employ either gradient boosting or transformer-based regression for the value ultimately predicted.
Tools: TensorFlow for deep learning, XGBoost for gradient boosting, OSM and possibly Zillow dataset.
Multimodal Learning, Dynamic Pricing Models, Explainable AI in pricing.
K. Kuppan et al., 2024.
C. Zhang, J. Sun, Y. Qi and X. Hu, “A survey of dynamic pricing: From the perspective of machine learning,” IEEE Access, vol. 8, pp. 187212–187228, 2020.
L. Wei, H. Zhou and J. Huang, “Multimodal Learning for Property Value Estimation,” in Proc. IEEE International Conference on Big Data, 2019.
While property valuation systems are black boxes and seem not open to scrutiny, they are not adopted in regulated real estate businesses.
Develop a computer vision model for property valuation that uses visual traits.
Develop a model that incorporates explainability (Grad-CAM, SHAP), that can substantiate their predictions.
How do consumers find meaning and difference between the formats of digital fashion end-products?
On what basis do consumers classify those end products?
Using labeled interior/exterior images, with appraisal values.
Train a regression-based CNN or ViT model.
Provide a visualization of what areas of the image most contributed to the prediction for the price.
Tools: PyTorch, SHAP, Grad-CAM, Label Studio (for annotation).
Explainable AI (XAI), Visual Interpretability in CV.
K. Kuppan et al., 2024.
R. R. Selvaraju et al., “Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization,” in IEEE ICCV, 2017.
S. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” in Advances in Neural Information Processing Systems (NeurIPS), 2017. (SHAP)
Real-time energy optimization of smart buildings could potentially infringe upon the user´s privacy as the data is stored in a centralized manner.
Build a privacy preserving AI based solution for smart energy management.
Use federated learning with differential privacy tech to protect the data.
Simulate multi-zone building with sensory inputs (HVAC, Lighting).
Application of FL (with local training updates) with DP-noise.
Tools: TensorFlow, Smart Building Simulator, PySyft, DP-SGD.
Smart Building AI, Differential Privacy (Dwork et al.), IoT energy optimization.