Selected work — production SaaS, marketplaces, and integration backends

Problem: Run several marketplace verticals on one shared, multi-tenant platform without forking the backend per vertical.

What I built: Built shared platform services on a TypeScript monorepo — authentication, Stripe Connect payments, and a multi-channel notification engine over Server-Sent Events — with background workers, the outbox pattern for consistent event delivery, and role-based access control on PostgreSQL / Drizzle and Redis.

Outcome: Shared backend infrastructure powering multiple marketplace verticals from a single codebase.

Problem: Give every business its own branded loyalty program with digital cards that update the moment a customer earns or redeems a reward.

What I built: Built a multi-tenant NestJS + Next.js platform with digital loyalty cards for Apple Wallet and Google Wallet, QR-based staff scanning, rewards and redemptions, and a partner API with separate authentication planes — all backed by queue-based background processing on BullMQ and Redis, with Paymob payments and WhatsApp Cloud API notifications.

Outcome: Production loyalty SaaS with wallet-issued cards, QR redemptions, and a partner API.

Problem: Let salons and coaches take bookings, payments, and payouts on their own branded space while keeping schedules and revenue consistent.

What I built: Built a multi-tenant NestJS + Next.js platform with an availability engine and staff scheduling, Stripe payments with invoices, refunds, commissions and payouts, revenue reporting, per-staff Google Calendar sync, and WhatsApp notifications — with a transactional outbox and background queues keeping side effects reliable.

Outcome: Live production booking platform serving real businesses at getbookra.com.

Problem: Take short-form video from idea to published across platforms without manual steps — reliably, even when AI and social APIs fail.

What I built: Engineered a NestJS + BullMQ pipeline covering script generation, rendering workflows, approval gates, multi-platform publishing, and performance analytics — hardened with circuit breakers, retry and dead-letter handling across AI/LLM and social integrations.

Outcome: Fully automated script-to-publish pipeline with circuit-breaker and dead-letter resilience.

Problem: Keep medical-record requests synchronized across law-firm case systems (Litify/Salesforce, Filevine) and Zendesk — reliably and auditably, with sensitive data protected.

What I built: Built a PHP 8.2 / Laravel 12 integration middleware with OAuth 2.0 token management, HMAC webhook verification, idempotent processing, bidirectional status synchronization, full audit trails, and PHI/PII redaction — queued and containerized with Docker.

Outcome: Reliable bidirectional sync covered by 290+ passing PHPUnit tests.

Problem: A portfolio is a passive page — an AI assistant helping a hiring manager can only scrape prose, with no structured way to ask "does this person have NestJS backend experience?" and get a typed answer.

What I built: Built a read-only MCP server exposing this site's public data — profile, skills, projects, case studies, backend experience — as 7 Zod-validated, AI-callable tools. One transport-agnostic tool layer backs both a stdio server (for local testing with Claude Desktop/Code or the MCP Inspector) and a stateless Streamable-HTTP endpoint at /api/mcp, so both stay in sync by construction.

Outcome: A working technical-branding demo, not a recruiter-facing feature — proof of AI/backend integration skill: typed tool contracts, input validation, and safe public-data exposure. Full write-up in docs/MCP_CASE_STUDY.md.

Problem: Get live tournament scoring on-air without manual data entry, even when upstream sports-data providers are flaky.

What I built: Built a Node.js bridge that polls sports-data providers, normalizes scoring data, and pushes it directly to the broadcast graphics platform over its real-time data API — with fault tolerance, a last-good cache, and rate-limit backoff.

Outcome: Powered live on-air graphics across multiple simultaneous tournaments.

Problem: Let client companies build and deploy conversational AI agents across chat, web, and phone — each grounded in its own knowledge and isolated per tenant.

What I built: Built the backend for a multi-tenant AI-agent platform: agent lifecycle (versioning, rollback, deploy-to-production), RAG retrieval on Vertex AI, voice agents over Twilio + Gemini Live via Python Google Cloud Functions, Stripe billing, and tool/webhook execution — with API-key auth, per-tenant isolation, and audit logging.

Outcome: Conversational agents serving multiple tenants across chat, web, and voice from one platform.

Problem: Turn dormant professional networks into a marketplace for warm introductions — with credits escrowed and automation that can't double-pay.

What I built: Built the marketplace backend: an escrowed credit ledger with idempotent, transactional claim/refund logic; a queue-based automation engine distributing LinkedIn connection jobs across a PhantomBuster account pool with circuit breakers and a dead-letter queue; webhook-driven contact enrichment via Wiza; and cron orchestration for distribution, expiry, and monitoring.

Outcome: Reliable credit operations — idempotent claims, circuit-breakered automation, and dead-letter recovery.