Aditya Sen Gupta AI Product Engineer & Systems Architect

The Problem

The business needed a premium e-commerce feel with localized shipping logic, but wanted to avoid the high recurring fees and rigid structures of platforms like Shopify.

The Architecture

Decoupled a mobile-first Next.js frontend from a normalized relational Supabase backend. Built a state-managed, merge-safe cart that pipes directly into a frictionless WhatsApp checkout, reducing drop-off. Pricing lives at the product level, stock is tracked per variant and per size, and district-aware logistics compute delivery charges for every district in Bangladesh.

• AI shopping concierge — a floating assistant built on the Vercel AI SDK and Gemini 2.5 Flash with live Supabase tool-calling: it answers real-time color, stock, and price questions from the actual inventory and can trigger the checkout flow directly from the conversation.
• Quota-resilient LLM orchestration — a fail-fast fallback chain (Flash → Flash-Lite), a server-side cooldown when quotas are exhausted, and hard per-message spend caps, so a free-tier 429 never becomes a customer-facing crash.
• Conversational WhatsApp checkout — the cart aggregates, formats, and hands the order to WhatsApp for confirmation and payment, removing multi-step checkout friction.
• Merge-safe persistent cart — a single cart contract shared across the storefront and admin previews, persisted to localStorage and reconciled across tabs so items are never lost.
• Spatial glassmorphism UI — a depth-layered hero and Framer Motion micro-interactions delivering the luxury feel the brand demanded, without sacrificing load performance.

The Impact

Zero platform fees, sub-second load times via Incremental Static Regeneration (ISR), and a custom admin command center — product/variant/size CRUD, an order pipeline with CSV export, review moderation, and inventory analytics — secured end-to-end with Row-Level Security (RLS).

The Problem

The original "The Whites Stories" Facebook page, which had grown to a massive 250k followers, was lost. The community needed a secure, independent home to prevent future platform deplatforming, alongside a targeted rebrand for the local audience.

The Architecture

Consolidated content, identity, community, and commerce into a single Next.js application. Built a Markdown-powered newsroom CMS, authenticated member accounts, and a secure merch storefront — one codebase, one auth system, one design language.

• Autonomous match-data pipeline — Vercel Cron endpoints sync fixtures and results from live football APIs into Postgres on a daily schedule, so matchday data stays current with zero manual entry.
• Rich-text editorial newsroom — Markdown-authored articles rendered through react-markdown with raw-HTML and line-break support, styled by Tailwind Typography; every story gets its own shareable slug with one-tap Facebook and X share intents.
• Fans Zone & Squad Builder — an interactive lineup builder that exports shareable matchday graphics as PNGs, generated entirely client-side, plus "Voices of the Madridistas," a moderated fan-review system.
• Server-rendered identity — cookie-based Supabase Auth sessions available inside React Server Components, gating member features and protected routes.
• Search-engine plumbing — automated sitemap and JSON-LD structured data with Google Publisher Center registration, turning the newsroom into an indexable publication rather than a feed.

The Impact

Reclaimed platform ownership from social media giants, creating a unified, app-like experience. The new, highly targeted "The Whites Bangladesh" community has already surged past 30k+ followers, proving the resilience of the brand.

The Architecture

An AI tutor leveraging Retrieval-Augmented Generation (RAG). Uses pgvector to store and semantically search O-Level physics marking schemes, allowing the Claude API to grade student answers and provide targeted, conceptually accurate feedback in real time.

marking schemes → embed → pgvector store
student answer  → semantic retrieval → top-k scheme points
                → Claude grading → targeted feedback

The Architecture

A specialized internal SaaS tool designed to automate boutique e-commerce operations. Leveraging Claude's advanced function calling, the agent will have direct read/write access to a Supabase database to autonomously draft inventory reports, flag low-stock variants, and generate localized marketing copy based on real-time sales velocity.

sales velocity → Claude agent (tool-use loop)
               ├─ read / write Supabase
               ├─ draft inventory reports
               ├─ flag low-stock variants
               └─ generate localized marketing copy