Projects
A selection of machine learning and AI engineering projects — from research prototypes to production deployments. Each project links to a deeper case study.
Novel federated conformal prediction method that mitigates hallucinations in federated fine-tuned vision-language models via abstention.
Impact: Reduced hallucination rate 13.05% → 3.79% (71% relative drop); useful-answer rate 76.63%; Precision@Commit 95.29%.
Attention-based forecasting of magnetic field trajectories combined with supervised contrastive learning for solar flare classification.
Impact: TSS = 0.75 on flare classification; outperformed persistence baseline on 60-step delta forecasts.
Imbalanced medical image classification using MobileViT v2 with dual-weighted learning — inverse square-root sampling + ENS loss — and Grad-CAM heatmaps for clinical interpretability.
Impact: 0.92 weighted F1 on an imbalanced skin lesion dataset.
Custom Temporal Attention mechanism in PyTorch for market regime prediction, with a differentiable Sharpe Ratio Loss that optimizes directly for risk-adjusted returns.
Impact: 15% improvement over MSE baselines in backtesting; Sharpe Ratio Loss outperforms standard regression objectives.
A framework for generating synthetic Q&A datasets using Gemini AI to benchmark semantic search models across diverse domains — eliminating the need for manual evaluation set curation.
Impact: Automated, domain-agnostic benchmarking framework for semantic search evaluation.
Open-source semantic image search supporting text→image, image→image, and text+image→image queries on the Myntra Fashion Product Dataset.
Impact: Real-time multi-modal retrieval across text and image modalities on a large fashion catalog.
Currency detection and classification for Nepali banknotes using InceptionV3 transfer learning, achieving 94% accuracy across 7+ denominations.
Impact: 94% accuracy on 7+ Nepali banknote classes. Updated in 2023 with additional denominations.