From complex requirements to production-ready software.

Retail businesses live on the checkout counter, the stock room and a live picture of what is moving where. Single-shop POS tools fall apart the moment a brand opens a second branch: sales, staff and inventory need to stay isolated per branch while management watches the whole business in real time — and every checkout has to cope with barcode scanners, concurrent sales and stock updates arriving from many locations at once.

• Design a multi-tenant, multi-branch data model so each organisation owns its branches, users and inventory behind strict tenant isolation.
• Own the stack end-to-end — NestJS with Prisma and PostgreSQL on the backend, React with Redux Toolkit on the frontend.
• Integrate hardware barcode scanners into the checkout flow so line items are added with a single scan and the sale keeps moving.
• Stream stock and sales changes over WebSockets so every branch sees live quantities against shared inventory.
• Move heavy work — sales aggregation, stock recalculation and report generation — onto BullMQ and Redis workers instead of blocking the API, cutting processing latency by roughly 40%.
• Add attribute-based access control so staff permissions are scoped per branch and per capability.
• Ship it like production: Docker orchestration, GitHub Actions CI/CD, an Nginx reverse proxy and automated database backups.

A scalable POS SaaS that runs multiple branches from one platform — barcode-driven checkout, real-time inventory sync and business reporting, with tenant isolation, queue-based background processing and a fully automated delivery pipeline.

• Multi-branch sales, stock and operations in one system
• Real-time inventory sync across branches via WebSockets
• ~40% lower processing latency with BullMQ + Redis
• ABAC permissions scoped per branch and role
• CI/CD, containerised deploys and automated backups

Generic uptime checks tell you an endpoint is down, but teams also need to understand degrading behaviour and get alerts that are actually worth reading. The project focused on turning raw API signals into meaningful, prioritised alerts.

• Build a monitoring backend that collects API metrics and request behaviour over time.
• Detect anomalies from the collected data instead of relying on a single static threshold.
• Integrate an LLM step to turn detected anomalies into readable, contextual alerts.
• Expose the results through an API for the dashboard to consume.
• Build a real-time React dashboard with Redux Toolkit for state and live updates.

A Node.js/Express monitoring service with anomaly detection and AI-generated alert summaries, surfaced through a Vite + React dashboard that visualises anomalies as they happen.

• Anomaly detection on API behaviour rather than binary uptime
• AI-assisted, human-readable alerting
• Real-time dashboard for monitoring at a glance
• Separated backend and frontend for independent scaling

Image search needs more than keyword matching: it has to understand visual similarity at scale while staying fast under real query load.

• Generate image embeddings with the OpenAI CLIP model.
• Store vectors in PostgreSQL using the pgvector extension.
• Add an HNSW index to keep similarity queries fast as the catalogue grows.
• Run embedding generation asynchronously so uploads stay responsive.
• Expose product CRUD and search through a typed FastAPI service.

A FastAPI backend that turns uploaded product images into vectors, indexes them for approximate nearest-neighbour search and returns visually similar products in real time.

• Visual, image-based product search
• Approximate nearest-neighbour matching with HNSW
• Asynchronous background embedding generation
• Typed, documented API with full CRUD