Shaurav Khadka

RedditPulse: Social-Media Intelligence Platform

Co-developed a seven-module NLP and LLM prototype over 1,989 filtered posts and 1,588 comments from 10 subreddits. The platform integrates preprocessing, NER, TF-IDF, sentiment modelling, zero-shot topic classification, sentence-transformer embeddings, FAISS retrieval, grounded insight generation, and a Streamlit dashboard. Retrieval evaluation reached Precision@5 = 0.68 and Recall@5 = 0.68.

Collaborative academic prototype.

• Python
• NLP
• RoBERTa
• RAG
• FAISS
• Streamlit

Production AI Reliability and Document Intelligence

Built a repeatable OCR-evaluation workflow for veterinary-insurance invoices across multiple OCR configurations, transformation rules, confidence scores, and error codes. Supported adversarial fraud-detection evaluation and produced recommendations on traceability, reproducibility, validation dependencies, latency, and cost.

Production-oriented AI/ML R&D internship work.

• Document Intelligence
• OCR
• AWS S3
• Azure Document Intelligence
• Python
• AI Reliability

Fine-Grained Vision and ROS2 Robot Deployment

Co-developed a pasta-recognition and vision-to-action robot pipeline. A ResNet50 model achieved 90.69% accuracy across 21 subclasses; separate three-class experiments reached 99.21%. Deployment-specific fine-tuning and augmentation improved robot-image accuracy from 2.38% to 95.24%, with predictions mapped to ROS2 navigation commands.

Collaborative computer-vision deployment project.

• ROS2
• Computer Vision
• PyTorch
• ResNet50
• Transfer Learning
• Sim2Real

Reinforcement-Learning Benchmark Suite

Implemented and compared agents across Taxi-v3, LunarLander-v3, AirRaidNoFrameskip-v4, and custom MountainCar-v0 wrappers. Frame skipping and frame stacking improved AirRaid PPO mean reward from 300 to 1,925. A momentum-position reward design achieved a 100% final MountainCar success rate.

• Reinforcement Learning
• PPO
• DQN
• Q-Learning
• Gymnasium
• Stable-Baselines3

Temporal GNN for Blockchain Fraud Detection

Modelled blockchain transactions as dynamic graphs and built a TGAT-style architecture with temporal encodings, attention-based message passing, and dual-task learning for fraud detection. Developed an end-to-end graph-learning pipeline and benchmarked temporal modelling against an XGBoost baseline.

• PyTorch
• NetworkX
• TGAT
• Temporal GNNs
• XGBoost
• Fraud Detection

LLM-Powered Financial Advisory and Multilingual AI Prototype

Built a retrieval-augmented decision-support prototype combining document retrieval, semantic search, vector embeddings, and context-aware LLM responses. Extended conversational workflows with multilingual language detection, translation support, and cross-lingual context management.

Research prototype. Not financial advice.

• LLMs
• RAG
• Semantic Search
• Vector Embeddings
• Multilingual NLP
• Conversational AI