Portfolio
A showcase of my machine learning projects.
A showcase of my machine learning projects.
● Performed time-series forecasting on NEPSE (Nepal Stock Exchange) data, leveraging statistical and machine learning models to predict market trends and evaluate predictive performance.
● Cleaned, processed, and transformed raw financial data (e.g. handling missing values, normalization, feature engineering) to create robust input features for forecasting models.
● Compared and validated multiple forecasting approaches (e.g. ARIMA, SARIMA, LSTM, Prophet, etc.), evaluated accuracy using metrics (RMSE, MAE, etc.), and provided insights into optimal model selection for stock index prediction.
View Code● Developed a comprehensive two-stage system: initial implementation with Inception-ResNet-v2, later enhanced with YOLOv8.
● Integrated state-of-the-art object detection with Tesseract OCR for robust license plate localization and text extraction.
● Achieved robust performance across diverse environmental conditions including varying lighting, angles, and image quality.
View Code● Developed the benchmark dataset for Nepali Sign Language (NSL) with 36 gesture classes and 1,500 samples per class.
● Fine-tuned MobileNetV2 and ResNet50 architectures achieving classification accuracies of 90.45% and 88.78% respectively.
● Demonstrated effectiveness of transfer learning and fine-tuning for underexplored sign languages in low-resource settings.
View Code● Fine-tuned Microsoft DialoGPT-medium on a synthetically generated dataset of 1,000 doctor-patient dialogues covering ten common diseases prevalent in rural Nepal to create an offline-capable medical conversational AI system.
● Designed a two-stage data generation and validation pipeline using Gemini 2.5 Pro and Claude 4 Sonnet, followed by expert medical review, ensuring accuracy, empathy, and contextual relevance of dialogues.
● Conducted quantitative (perplexity, loss, token-level F1) and qualitative evaluations by healthcare professionals, demonstrating the model's ability to produce coherent, medically appropriate, and empathetic responses for low-resource healthcare settings.
View Code● Implemented an end-to-end evaluation pipeline for Amazon Bedrock models (multi-region support) with performance (latency, throughput, time-to-first-token) and quality metrics (helpfulness, faithfulness, completeness, coherence, etc.).
● Used LangChain orchestration and an LLM-as-a-judge approach for automated, multi-dimensional quality and responsible-AI assessments (harmfulness, bias, refusal appropriateness), plus async processing for concurrent benchmarking.
● Produced structured evaluation outputs and human-readable reports to compare models, tune prompts, and inform safe deployment decisions.
View Code● Engineered an automated research agent pipeline that uses Amazon Strands Agents for domain-specific data gathering, filtering, and summarization, reducing manual research time by a significant margin.
● Implemented a multi-agent architecture combining “Agent Loop,” “Agent-as-Tool,” “Web Search,” and “Extraction” modules to autonomously perform requirements analysis, competitive landscape research, and effort estimation.
● Incorporated iterative feedback loops and inter-agent coordination to refine project scoping, synthesize comparative analyses, and output actionable estimations and insights.
View Code● Conversational AI systems (Voice and Chat) for healthcare applications.
● Automated foundational models evaluation pipelines with comprehensive performance reporting.
● AI agent frameworks and Model Context Protocol (MCP) clients and servers.
● AI-powered voice interviewing systems.
● Sophisticated job-candidate matching algorithms.
● Contextual search systems leveraging embedding models with multi-service integration.