Quill & Pigeon

Quill & Pigeon helps customers discover, personalize, schedule, and send handwritten cards created by independent New England artists.

I joined as a Full-Stack Software Engineer Co-op and worked across the customer-facing Next.js platform, Medusa commerce services, Stripe subscriptions and payments, AWS Lambda applications, search infrastructure, transactional email, shipping workflows, and context-aware AI agents.

I worked across the Next.js platform, Medusa commerce services, Stripe subscriptions and payments, AWS Lambda applications, search infrastructure, transactional email, shipping workflows, and context-aware AI agents.

Users supported

500+Users supported

Transactions processed

1,000+Transactions processed

Faster page loads and processing

50%Faster page loads and processing

Accessible customer experience

WCAG 2.2 AAAccessible customer experience

Built accessible customer workflows across product discovery, personalization, recipients, reminders, subscriptions, checkout, and card delivery using Next.js, React, TypeScript, Prisma, Zod, Medusa, and PostgreSQL.

These flows connected the storefront to commerce and fulfillment services so customers could discover cards, personalize them, manage recipients and important-date reminders, subscribe, check out, and track delivery — all with an accessible, WCAG 2.2 AA experience.

Medusa did not natively support Quill & Pigeon's subscription card-credit model, so I extended its commerce and payment architecture to support it. Customers received card credits based on their subscription tier and could redeem those credits to receive included or free cards during checkout.

Extended Medusa's commerce and payment architecture to support subscription card credits, allowing customers to redeem included cards during checkout while preserving Stripe-based payment workflows for standard purchases and remaining balances.

• •Customers received card credits based on subscription tier.
• •Credits could be used to receive included or free cards.
• •Credit redemption was integrated directly into checkout.
• •Custom Medusa functionality backed the credit model — it is not a native Medusa capability.
• •Custom payment-provider behavior connected Stripe payment workflows and credit redemption.
• •Checkout logic distinguished between credit-funded and standard paid purchases.
• •Subscription lifecycle events updated customer-credit availability.

Custom Stripe payment provider

Built custom payment-provider behavior that integrated subscription card-credit redemption with Stripe-backed payment and checkout workflows. The design keeps three concerns clearly separated: Stripe handled payment processing, the custom provider handled subscription-card credit redemption, and Medusa handled order creation. Card credits are an application-level entitlement — not cryptocurrency or a cash equivalent.

Built an MCP server that supplied AI agents with authenticated customer context, including subscription tier, available card credits, and relevant account information, allowing responses and recommendations to reflect the customer's actual product state.

• •The AI could provide account-aware answers.
• •Recommendations could consider the customer's subscription.
• •The agent could determine whether the customer had credits available.
• •Context did not need to be manually copied into each prompt.
• •Context was exposed through controlled MCP tools.
• •The AI did not receive unrestricted database access.

Customer context was exposed through controlled MCP tools rather than direct, unrestricted access to the application database.

Built AI-powered product and support workflows using the OpenAI API and Claude API, with provider failover and customer context supplied through the MCP server. Instrumented these AI workflows with Langfuse for prompt management, cost visibility, tracing, debugging, and understanding agent decisions — Langfuse provides observability around the workflows rather than making AI decisions itself.

Integrated Meilisearch to support fast product discovery and search across the commerce experience, keeping product lookups responsive as the catalog grew.

Implemented email conversation chaining so AI workflows received the relevant conversation history before reacting to inbound messages.

Built conversation-aware inbound email handling that reconstructed the relevant message thread before invoking the AI workflow. Requests the agent could not safely or confidently resolve were escalated to the owners for human follow-up.

• •Inbound emails were associated with their existing conversation.
• •Prior messages were reconstructed or retrieved.
• •The AI received the relevant thread context.
• •The AI attempted to handle the request using the full conversation.
• •Requests that could not be handled safely or confidently were escalated to the business owners.
• •Human owners could continue the existing conversation.

Branded transactional email templates were built with React Email and connected to automated customer workflows, covering communication such as event reminders, order emails, and subscription updates alongside conversation-aware inbound handling.

Prisma and Kysely served different architectural needs and were not used interchangeably in the same runtime.

Integrated USPS shipping estimates into the checkout and fulfillment experience so customers could see more accurate delivery expectations. These are estimated delivery windows, not guaranteed delivery dates.

Configured GitHub Actions workflows with AWS OIDC authentication and used act to run and debug CI workflows locally before pushing changes. act is a local GitHub Actions runner — not an AWS service, a replacement for GitHub Actions, or a production deployment platform.

• •Reduced reliance on long-lived AWS credentials through OIDC authentication.
• •Improved confidence when modifying workflows.
• •Reduced iteration time when debugging GitHub Actions with act.
• •Supported automated testing and AWS deployments.

Validated API and application boundaries with Zod and tested service integrations using Bruno and Postman. Maintained consistent formatting and review quality using Prettier and automated repository checks.