Playground
Nisum Yonghang / AI Engineer

Developing agentic applications, custom models, and RAG pipelines.

I am a Software Developer and AI/ML Specialist building production-grade full-stack architectures, Large Language Model pipelines, and real-time computer vision systems.

Retrieve Projects Open Playground
System.Overview

Personal Profile

Profile.txt

Nisum Yonghang

LLM / AI / ML Enthusiastic & Engineer

Currently pursuing CSIT at St. Xavier's College, Kathmandu. Passionate about Retrieval-Augmented Generation (RAG), multi-agent systems, parameter-efficient fine-tuning (QLoRA), and deep learning. I bridge the gap between complex machine learning theory and production-grade web systems.

CSIT Student AI Researcher Full-Stack Dev Leader Sports Coord
Contact.json

Quick Links

Education.md

Studies

2022 - Present

St. Xavier’s College

B.Sc. CSIT | GPA: 3.65

2019 - 2021

Little Angel’s College

Science (Biology) | GPA: 3.22

Certificates.log

Credentials

AI_ENGINEER

AI Engineer Core Track

Udemy | Issued Jun 2026

ID: UC-f3743c60-4219-4f83-a722-e424630da4bb
DS_ML

Data Science & ML

Broadway Infosys | Issued Aug 2025

Credential: B50011000
AWS_CCNA

CCNA Routing & Switching | AWS Academy

St. Xavier's College | 2025

System.Projects

Featured Repositories

A curated collection of neural pipelines, RAG backends, and full-stack software.

AI_AGENT

HOT_DEALS — Multi-Agent AI System

Autonomous multi-agent system designed to estimate the fair value of live product listings and automatically identify high-value bargain deals in real-time.

Python LLMs LangChain RAG Multi-Agent
DETECTION

Vehicle Detection and Counting System

Real-time computer vision system that tracks, counts, and analyzes vehicles crossing specific boundaries using state-of-the-art YOLOv11 architecture.

Python YOLOv11 OpenCV PyTorch NumPy
CLASSIFY

Tea Leaf Disease Classification

Image classification network trained on real datasets in Bhadrapur to diagnose plant pathogens, utilizing Grad-CAM mapping for visual model explainability.

Python PyTorch OpenCV Grad-CAM Transfer-Learning
GESTURE

Hand Sign Detection System

Real-time computer vision framework using convolutional neural networks to segment and interpret dynamic hand gestures for human-computer interaction.

Python PyTorch OpenCV NumPy CNN
MEDICAL

Pneumonia Detection in Chest X-Rays

Medical imaging diagnostic model based on a custom CNN architecture to detect pneumonia with visual interpretability layers to assist clinical decision making.

Python PyTorch Scikit-Learn Matplotlib Grad-CAM
System.Experience

Professional Timeline

Full Stack Developer

Sunkoshi Manpower Service Pvt. Ltd.
Apr 2026 – July 2026 | Freelance

Developed and deployed a production-ready full-stack website and web application for a manpower recruitment agency, coordinating the lifecycle from requirement analysis to final production deployment.

  • Collaborated directly with corporate managers to gather business requirements and implement them as scalable technical codebases.
  • Built a responsive, content-rich web client with streamlined job application and tracking capabilities.

Software Developer

Kashvi Education Institute and Research Center
Jan 2026 – Apr 2026 | Contract

Designed and built a modular Educational Management System to optimize administrative processes and student records management.

  • Improved administrative efficiency and database throughput by refactoring data access layers.
  • Collaborated directly with stakeholders to review operational bottlenecks and architect solutions for multi-institute scaling.
System.Toolkit

Tokenizer Frameworks

A list of libraries and programming languages Nisum utilizes.

Languages

Python Javascript C++ Java C PHP

AI & LLMs

Large Language Models (LLMs) RAG Multi-Agent Systems Agentic AI Development LangChain LiteLLM QLoRA / LoRA Fine-Tuning

ML & CV

Deep Learning PyTorch Computer Vision OpenCV YOLOv11 Object Tracking CNNs Transfer Learning Data Analysis

Tools

AWS Git & GitHub MySQL Jupyter Notebooks Postman Label Studio VS Code
System.Playground

AI Agent Playground

Tweak system settings, models, and temperature levels to interact with Nisum's profile.

Playground Custom Agent Console
System Instructions
System:

Console initialized. Adjust the settings in the right sidebar to test different response types. Click any command button on the right to start querying.

HOT_DEALS — Multi-Agent AI System

Role: Lead AI System Designer
Tech Stack: Python, LangChain, LiteLLM, RAG, Gradio

The Challenge

Analyzing live product listings across disparate sources to identify bargains is a human-intensive task requiring data extraction, historical value analysis, quality verification, and bargaining strategy.

My Solution

I architected an autonomous multi-agent system where specialized AI agents cooperate to solve this workflow:

  • Information Retrieval Agent: Scrapes, normalizes, and filters live product data.
  • Pricing Analysis Agent: Evaluates historical price trends and computes standard market values.
  • Reasoning Agent: Compares listings against cached models and reasons whether a listing is a true bargain.
  • Decision Agent: Signs off on high-confidence bargains and initiates automated alerts.

I integrated Retrieval-Augmented Generation (RAG) to feed contextual market dynamics into the reasoning pipeline and deployed a live Gradio Dashboard showing active bargain scoring and alert triggers.

Vehicle Detection & Counting System

Role: CV Developer
Tech Stack: Python, YOLOv11, OpenCV, PyTorch, NumPy

The Challenge

Traffic monitoring systems need to be lightweight and accurate, running on low-resource hardware at intersections to track specific vehicles and count flows in real time.

My Solution

I engineered an end-to-end computer vision pipeline using YOLOv11 for high-frequency object detection:

  • Integrated custom OpenCV bounding-box intersection calculations to detect vehicles crossing defined virtual lines.
  • Built mathematical centroid tracking loops assigning unique IDs to every vehicle to prevent double counting.
  • Handled motion vectors to track multi-directional lanes and export summary logs for transportation analysts.

Tea Leaf Disease Classification

Role: Machine Learning Engineer
Tech Stack: Python, PyTorch, torchvision, OpenCV, Grad-CAM

The Challenge

Farmers in rural tea estates like Jhapa face difficulties diagnosing plant pathogens early, resulting in reduced crop yield. The diagnosis needs to be explainable so field technicians trust the model.

My Solution

Built and validated an image classification neural network based on convolutional layers:

  • Collected and preprocessed a specialized dataset of healthy vs diseased leaves directly in Sunkoshi Tea Estate.
  • Optimized transfer learning using PyTorch models to train under limited dataset size constraints.
  • Integrated Grad-CAM visualization layers to overlay color-activation maps, highlighting exactly where the model detects disease indicators (e.g. leaf spots, blight outlines).

Hand Sign Detection System

Role: Lead Vision Researcher
Tech Stack: Python, PyTorch, OpenCV, NumPy

The Challenge

Creating an interactive framework that maps dynamic hand poses in real time to digital interface controls requires high frame rates, robustness under ambient lighting, and high classification accuracy.

My Solution

Designed a custom convolutional neural network (CNN) trained on segmented gesture datasets:

  • Developed pre-processing scripts utilizing skin-tone masking and morphology filters to isolate hand targets from background clutter.
  • Trained custom PyTorch models to recognize alphabet gestures.
  • Integrated tracking metrics using NumPy distance algorithms, enabling direct interface scrolling and clicks mapped directly to specific hand gestures.

Pneumonia Detection in Chest X-Rays

Role: Deep Learning Specialist
Tech Stack: Python, PyTorch, Scikit-Learn, Matplotlib, Grad-CAM

The Challenge

Chest X-rays require expert radiological interpretation. Early screening tools in underserved clinics can support diagnostic speed if they achieve high sensitivity and explainability.

My Solution

Engineered a convolutional neural network designed specifically for medical image anomalies:

  • Applied advanced data augmentation to prevent overfitting on clinical-source datasets.
  • Trained models in PyTorch and calculated validation thresholds utilizing detailed ROC curves and precision-recall metrics.
  • Mapped model decisions visually using Grad-CAM, enabling physicians to trace activation weights directly onto pulmonary densities.