Leave Management Portal
Leave Management Portal
A web-based portal developed to manage employee leave requests and approvals for 150+ employees at Fore Solutions. The system was designed to replace manual HR processes and improve transparency.
Key Features:
Role-based workflow (Employee → Manager → Director → HR)
Leave balance tracking with auto-deduction
Excludes weekends & holidays in leave calculation
Manager/Director/HR dashboards with approval/rejection workflow
Tech Stack:
Django | Python | Bootstrap | MySQL
My Role:
Designed the database models (Users, Leave, LeaveBalance)
Implemented leave approval routing logic (manager → director → HR)
Built HR summary dashboards with filters & modals
Integrated holiday/weekend exclusion logic
Impact:
Automated leave workflow → saved HR team hours weekly
Improved accuracy in leave balance tracking
Easy access dashboards for managers & directors
A web-based application designed to automate expense tracking, approval, and reimbursement for employees. The goal was to reduce manual paperwork, improve approval workflows, and give management better visibility on company spending.
Key Features:
Employees can submit expense claims with attachments (bills/receipts)
Multi-level approval workflow (Manager → Finance → Director)
Auto-categorization of expenses (Travel, Food, Office, etc.)
Dashboard for employees (status tracking) and Finance team (budget overview)
Export reports in Excel/PDF for audits
Tech Stack:
Django | Python | Bootstrap | MySQL
Optional: REST APIs for report export
My Role:
Designed database models for Expenses, Categories, Approvals
Built workflow automation logic for approvals
Integrated file upload feature (bills/receipts)
Developed dashboard UI for employees & finance team
Impact:
Reduced manual expense approval time by 40%
Provided real-time visibility into company expenses
Improved accuracy in financial records and reduced errors
Developed an AI-powered chatbot capable of answering user queries from uploaded PDF documents using Retrieval-Augmented Generation (RAG).
The system integrates document ingestion, vector database, and local LLM inference, ensuring fast, secure, and cost-free deployment.
Multi-PDF Support – Upload and query multiple PDFs.
Semantic Search – Uses embeddings for accurate information retrieval.
RAG Architecture – Combines vector search with LLM reasoning.
Persistent Vector Store – Documents stored in ChromaDB for reuse.
Local Inference – Runs with Ollama (Llama3/Qwen) for privacy & zero cost.
FastAPI Backend – Provides REST API for frontend/web apps.
Cited Answers – Responses reference PDF page numbers for transparency.
LLM: Ollama (llama3.1, qwen2.5)
Embeddings: HuggingFace (intfloat/e5-small-v2)
Vector DB: ChromaDB
Backend: FastAPI
PDF Parsing: PyPDF
Deployment: Docker + Uvicorn
Upload PDFs → Extract text with PyPDF.
Chunk & Embed → Convert text into embeddings with HuggingFace.
Store → Save chunks + metadata in ChromaDB.
Query → User asks a question.
Retrieve → Relevant chunks retrieved via similarity search.
Generate Answer → LLM responds using retrieved context, citing sources.
Research assistant for academic PDFs
Company policy Q&A bot
Legal document search assistant
Financial report summarizer
A custom Python-based SSH client created to remotely connect, execute commands, and manage servers. This tool automated routine admin tasks and simplified DevOps workflows by replacing manual SSH logins with a scriptable solution.
Key Features:
Secure authentication using username/password & SSH keys
Execute single or multiple commands across remote servers
Support for parallel connections to manage multiple hosts
Integrated error handling & logging of executed commands
Extensible for automation tasks (file transfer, system checks)
Tech Stack / Libraries:
Python | Paramiko (SSH library) | Logging | ConfigParser
My Role:
Developed the core SSH logic using Paramiko
Added support for multiple server connections from a config file
Implemented logging system to store command outputs/errors
Packaged the script for reusability by other team members
Impact:
Reduced manual effort for server administration
Allowed faster troubleshooting and configuration management
Improved productivity by automating repetitive SSH tasks
A real-time object detection system deployed on NVIDIA Jetson devices to process video streams from CCTV cameras and drones. The project focused on enabling low-latency detection at the edge for security and surveillance use cases.
Key Features:
Real-time video feed capture from CCTV/Drone cameras
Deployed YOLOv5/SSD models optimized for Jetson (TensorRT acceleration)
Detected multiple classes: people, vehicles, and objects of interest
Edge-side inference → no need to send raw video to cloud (privacy & speed)
Alert generation (optional: email/SMS) when object detected
Tech Stack / Tools:
NVIDIA Jetson (Nano / Xavier / Orin)
Python | OpenCV | TensorRT | DeepStream SDK | YOLOv5/SSD models
My Role:
Preprocessed datasets & trained detection model
Optimized inference pipeline for Jetson hardware (GPU acceleration)
Integrated OpenCV pipeline to capture frames from CCTV/Drone
Deployed real-time detection + alert notifications
Impact:
Achieved real-time inference (~20–30 FPS) on Jetson edge devices
Reduced bandwidth usage by processing locally instead of sending to cloud
Useful for surveillance, traffic monitoring, and security automation
As part of a large-scale AI infrastructure initiative, I successfully set up and configured a BasePOD cluster at Thapar University, Patiala, in collaboration with NVIDIA and a dedicated team. The deployment includes:
6 NVIDIA DGX H100 systems
4 NVIDIA DGX A100 systems
This project delivers petaflop-scale AI performance, enabling advanced research, large-scale AI model training, and HPC (High Performance Computing) workloads for the academic and research community.
Establish a robust AI supercomputing cluster for students, researchers, and faculty.
Integrate state-of-the-art NVIDIA DGX hardware for both AI training and inference.
Ensure scalability, security, and optimized resource utilization across all nodes.
Provide containerized and isolated environments for multiple users.
Enable seamless integration with frameworks like PyTorch, TensorFlow, RAPIDS, and distributed training libraries.
6 × NVIDIA DGX H100 (8× H100 GPUs each, NVLink, 80GB HBM3 per GPU)
4 × NVIDIA DGX A100 (8× A100 GPUs each, NVLink, 80GB HBM2e per GPU)
High-speed networking: NVIDIA Quantum-2 InfiniBand (400 Gbps), Ethernet switches
Storage: High-throughput parallel storage (NFS/GPFS/NetApp)
Management nodes: For orchestration, monitoring, and logging
Base Command Manager (BCM) for DGX management
NVIDIA GPU Cloud (NGC) for optimized AI containers
Container Orchestration: Docker + Kubernetes (or Slurm for HPC scheduling)
Monitoring & Logging: Prometheus, Grafana, DCGM
Security: SSH hardening, VPN access, network segmentation
JupyterHub/JupyterLab for interactive workloads
Pre-installed Deep Learning Frameworks: PyTorch, TensorFlow, MXNet, RAPIDS
Distributed training setup: NCCL, Horovod, DeepSpeed
Isolated user containers with GPU allocation policies
Designed the cluster topology interconnecting DGX H100 and A100 nodes with the help of NVIDIA Team
Deployed and configured Base Command Manager for orchestration
Set up multi-user containerized environments (Docker + Kubernetes)
Configured networking & InfiniBand fabric for low-latency GPU communication
Implemented monitoring dashboards (Grafana + DCGM) for GPU utilization and health
Ensured secure user access with authentication and role-based permissions
Delivered training-ready environments for 100+ researchers and students
This project was executed with strong teamwork and support:
NVIDIA engineers → Provided direct assistance and best practices for BasePOD deployment
Gurpreet → Handled critical networking setup and configuration (InfiniBand, Ethernet)
Shubam & Sahil→ Assisted in system administration, storage integration, and testing
My role → Led the cluster design, orchestration, monitoring, and deployment
Delivered a petascale AI infrastructure for cutting-edge research at Thapar University
Enabled large-scale AI model training (LLMs, Vision Transformers, GNNs)
Provided multi-user GPU environments with isolation and fairness
Created a future-ready cluster scalable for additional DGX systems
Fostered collaboration between university, NVIDIA, and research teams
At Amity University, Noida, I co-led the deployment of a Kubernetes-based AI cluster using:
2 × NVIDIA DGX H100 systems (latest Hopper architecture GPUs)
Dell Scheduler for workload management
In addition to the infrastructure setup, our team conducted an NVIDIA training workshop for faculty and students, covering AI infrastructure usage, containerization, and distributed training workflows.
Deploy a scalable Kubernetes-based AI cluster for academic research.
Integrate Dell Scheduler for efficient GPU scheduling and resource allocation.
Provide a multi-user containerized environment with GPU isolation.
Deliver hands-on training sessions to educate researchers on AI cluster usage.
2 × NVIDIA DGX H100 (8× H100 GPUs each, NVLink, 80GB HBM3 per GPU)
Dell Scheduler nodes for job management
Networking: High-speed Ethernet fabric
Kubernetes (K8s) for container orchestration
Dell Scheduler integrated with K8s for job queuing
NVIDIA GPU Operator for Kubernetes GPU workloads
NGC Containers (NVIDIA-optimized AI/ML frameworks)
Prometheus + Grafana for monitoring
JupyterHub for interactive workloads
Set up and configured the Kubernetes cluster across DGX nodes
Integrated Dell Scheduler with K8s for workload scheduling
Installed and tuned NVIDIA GPU Operator for GPU-aware workloads
Implemented resource quotas, RBAC, and multi-user access policies
Designed monitoring dashboards for GPU utilization and health
Prepared hands-on labs and demos for the NVIDIA training session
Shubam → Partnered with me for cluster setup, testing, and troubleshooting
Amity University IT Team → Provided physical infra support and networking assistance
NVIDIA → Assisted with container optimizations and provided training resources
Training Delivery:
Lalit, myself , and Shubam delivered NVIDIA training sessions on:
Using DGX systems for AI workloads
Running jobs with Dell Scheduler + Kubernetes
Best practices for multi-user GPU environments
Delivered a production-ready Kubernetes AI cluster on 2× DGX H100
Enabled researchers and students to run distributed training workflows
Improved job efficiency with Dell Scheduler integration
Educated faculty and 100+ students through hands-on NVIDIA training workshops
Established a scalable foundation for future DGX expansions
Successfully completed the installation and configuration of NVIDIA L40 GPUs in a Dell PowerEdge server at World College of Technology and Management (WCTM). The setup was prepared to support AI/ML and high-performance computing workloads.
Installation Tasks:
Installed NVIDIA L40 GPUs in Dell PowerEdge server chassis.
Configured BIOS/PCIe settings, power, and cooling for optimal GPU performance.
Installed and verified NVIDIA drivers and CUDA Toolkit.
Integrated Docker + NVIDIA Container Toolkit for GPU-enabled containers.
Performed benchmark tests to validate GPU functionality.
Tech Stack & Tools:
Hardware: Dell PowerEdge Server, NVIDIA L40 GPU
Software: Ubuntu Linux, NVIDIA Driver, CUDA, Docker
My Role:
Led the end-to-end installation process – hardware setup, driver installation, and environment configuration.
Verified stability and performance of GPUs post-installation.
Collaborated with Utkarsh Bharti for testing and server configuration support.
Outcome:
Delivered a fully operational Dell server with NVIDIA L40 GPUs ready for AI/ML workloads.
Ensured proper GPU integration and validation for future training and research use.