Selected Work
3D Indoor Scene Synthesis
Agentic AI optimization loop with RL fine-tuning so pretrained 3D scene generators satisfy functional constraints. Reduced scene collisions from 81% → 62% and out-of-boundary objects from 5% → 3%. Presented at CVPR 2026, Denver.
RL for Diffusion Post-Training
Novel RL-based post-training method for diffusion models that improves reward alignment while preserving image quality and diversity. Built automated evaluation pipelines for reward modeling, fidelity checks, and diversity metrics.
Human Action Recognition
VideoMAE-based model with 95%+ accuracy on Nepali video datasets. Log-linear sequence scaling for efficient long-video inference. Built with Khalti.
Fashion Virtual Try-On
Fine-tuned IDMD diffusion models for 20× faster virtual try-on without fidelity loss. Evaluated CatVTON, IDM-VTON, and OOTDiffusion pipelines. KU Hackfest AI Track Winner.
Olympic Sentiment Analysis
FastAPI + React web app with custom BERT model for real-time sentiment analysis of Olympic social media posts. GeeksForGeeks Hackathon runner-up.
Adaptive Traffic Light System
Intelligent traffic control using ESP32, Arduino, and computer vision for real-time vehicle detection and classification. Automated signal timing based on live traffic counts.
VLM Deepfake Detection
Ultra-lightweight VLM system for real vs. AI-generated image classification. 98.7% on seen generators, 87.3% on unseen — strong generalization for a deployable solution.
RAG Grading Agent
End-to-end RAG pipeline for automated exam grading in Kathmandu Valley schools. Document retrieval, LLM reasoning, prompt evaluation, and workflow automation.
