My Projects

A personal productivity app (tasks, habits, notes, focus timer) built with React, Go, and Postgres — deployed on a production Kubernetes cluster with GitOps delivery, full-stack observability, and automated CI/CD on AWS.

• Provisioned the complete AWS environment (VPC, subnets, security groups, IAM) using Terraform in ~50 seconds, replacing a multi-hour manual setup with a single repeatable command.

• Automated node configuration (swap, kernel modules, sysctl tuning) with Ansible in ~65s and bootstrapped a 2-node RKE Kubernetes cluster with Canal CNI and NGINX Ingress in ~2m 25s — blank AWS account to fully operational cluster in under 10 minutes.

• Eliminated configuration drift across all environments by authoring Kubernetes manifests for 3 application tiers with ConfigMaps, Secrets, and a Postgres StatefulSet backed by persistent storage — replacing ad-hoc kubectl changes with fully declarative, version-controlled configuration.

• Designed a GitHub Actions CI/CD pipeline that builds Docker images tagged with commit SHA and auto-updates deployment manifests, achieving zero-effort releases with consistent sub-3-minute deployments.

• Implemented GitOps delivery with ArgoCD, auto-syncing all 3-tier manifests on every commit with a full audit trail and zero manual kubectl interventions in production.

• Deployed a full observability stack (Prometheus, Grafana, Loki, Alertmanager) covering cluster metrics, pod logs, and application traces — with API latency dashboards and automated email alerting.

• Documented the complete system architecture using Mermaid diagrams covering AWS infrastructure, cluster topology, CI/CD pipeline, GitOps workflow, and the observability stack.

• Hardened cluster networking with restrictive AWS Security Groups, allowlisting only required ports (etcd 2379–2380, kubelet 10250, Canal VXLAN 8472 UDP) and enforcing least-privilege IAM throughout.

A containerized Next.js/Postgres application on AWS EC2 with a fully automated Jenkins CI/CD pipeline, zero-touch rollback, and production resilience validated through structured incident response drills.

• Engineered a 4-stage Jenkins CI/CD pipeline (build, test, deploy, verify) with consistent sub-2m30s deployments and a fully automated, zero-touch release track.

• Implemented automated rollback using versioned Docker image tags and a last-known-good marker file — health-check-triggered recovery with no manual intervention, ensuring no failed deployment ever reaches production.

• Ran structured incident response drills across three failure scenarios: ~5s crash recovery (SIGTERM), ~20s database outage with graceful 503 degradation, and ~116s full EC2 reboot recovery — all self-healing, zero manual intervention.

• Secured Postgres credentials via Jenkins credential binding, keeping secrets out of source control and enforcing least-privilege access across all deployment stages.

• Documented findings through RCA reports and operational runbooks.

Tech: Next.js, Postgres, Docker Compose, Jenkins, Nginx, AWS EC2, Prisma, Certbot SSL

A production resume builder serving real users, running as a 2-app Next.js monorepo — containerized with optimized Docker builds and automated deployment pipelines on AWS.

• Cut production Docker image size by ~90% by authoring multi-stage Dockerfiles with Turborepo pruning, shipping only the standalone Next.js bundle in the final stage.

• Eliminated race-condition failures between 2 interdependent services using Docker Compose with healthcheck-gated startup ordering, isolated bridge networking, and environment-variable injection — stable, reproducible deploys across all environments.

• Reduced per-deploy asset pipeline time by ~40% with Python and Bash automation scripts for image optimization and asset processing, removing manual steps from the release workflow.

• Maintained architecture documentation, deployment runbooks, and component SOPs, reducing onboarding time and establishing a reliable reference for the team.

A deep learning system that translates Romanized and Devanagari Nepali queries into English and retrieves relevant news articles — built with a team of 3.

• Designed an end-to-end NLP pipeline combining a fine-tuned mBART transformer for translation with the BM25 ranking algorithm for probabilistic document retrieval.

• Fine-tuned mBART on a Nepali–English parallel corpus, achieving a BLEU score of 39.84 and BERT score of 0.9613 — comparable to larger commercial models.

• Enabled users to search English news content using informal, chat-style Nepali inputs.

An HTTP server built from scratch using Node.js and raw TCP sockets — no frameworks, no abstractions, just the protocol.

• Implemented the full HTTP/1.1 request-response cycle manually: parsing headers, routing requests, and serving responses over raw TCP.

• Added gzip compression, structured error handling, and asynchronous file serving.

• The kind of project that changes how you think about every web server you touch after.

A configuration-driven REST API integrating OAuth 2.0 authentication and multiple payment gateways for third-party login and digital transactions.

• Implemented OAuth 2.0 via Laravel Socialite for secure third-party authentication.

• Integrated eSewa and Khalti payment gateways with test APIs to simulate end-to-end transaction workflows.

• Designed an API-first architecture for straightforward integration with web and mobile clients.

A full-featured content management system built with the MERN stack — JWT authentication, role-based access, and complete CRUD for posts and users.

• Implemented JWT-based auth with email verification and password recovery via Nodemailer.

• Used JOI for request validation and Redux for global state management.

• Designed a responsive UI with Tailwind CSS and Flowbite-React, deployed on Render with Firebase for media storage.