Mirza Shahbaz Ali Baig

Businesses in the service industry rely on fragmented tools — separate systems for scheduling, payments, client management, and marketing — with no unified platform built for their workflow. Existing solutions are either too generic or too expensive to white-label at scale.

Architecting a full multi-tenant B2B SaaS platform from the ground up. Each tenant is fully isolated at the data layer via a tenant-scoped Prisma middleware. Built a plugin-based PayableHandler registry with a central PaymentOrchestrator owning all Stripe Connect logic — handling packages, gift cards, memberships, and wallet credits under a single payments interface. A trigger-filter-action Automation engine (similar to Zapier Flows) was designed for internal approval, with BullMQ workers processing async jobs. Cloudflare R2 handles multi-tenant media with signed URL uploads. Resend powers transactional email with per-tenant custom domain verification.

Core modules — scheduling, payments, client CRM, and automation — actively in development. Architecture reviewed and approved for production scale. Targeting feature parity with leading vertical SaaS competitors.

An educational board serving 15,000–20,000 students needed real-time WhatsApp communication — results, alerts, announcements — with zero message loss and a live agent dashboard.

Designed a 4-phase webhook → queue → WebSocket architecture on Meta Business API directly (skipping Twilio to eliminate per-message costs at scale). Inbound webhooks validate HMAC-SHA256 signatures and enqueue instantly to BullMQ — returning HTTP 200 before any processing to prevent Meta retry storms. Workers handle idempotency, persistence, and Redis Pub/Sub. A Socket.IO gateway routes live events to agents by room.

Message visibility latency: 5,000ms → under 1,000ms. Polling load reduced by 92%. Live in production serving 15,000+ students.

Most AI chatbot tutorials produce demos that fall apart under real use — no persistent conversation state, no reliable tool-calling, and no path from "notebook script" to something actually deployable.

Built a production-structured agentic chatbot using LangGraph to model the agent as an explicit state machine rather than a single prompt-and-response loop, giving it controllable, inspectable reasoning steps instead of a black box. The agent has real tool-calling access — live web search via Tavily, academic paper lookup via arXiv, and Wikipedia retrieval — routed through LangChain, with Groq and OpenRouter powering low-latency inference. FastAPI exposes the agent as a proper backend service, paired with a Next.js frontend for the primary chat interface. Containerized both services with multi-stage Docker builds and deployed to AWS EC2 behind an Nginx reverse proxy with a dedicated subdomain and SSL.

A reusable, tool-augmented agentic foundation — not a demo — live in production and reachable at a public URL, with a clean separation between the reasoning layer (LangGraph agent), the API layer (FastAPI), and the interface (Next.js), so any piece can be swapped or extended independently for future client work.

Manual code review is slow and inconsistent — small teams and solo developers often push PRs with no second pair of eyes, letting bugs, style issues, and security gaps slip into main branches unreviewed.

Built a full AI-powered PR review platform that connects directly to GitHub via webhooks. Every pull request triggers an event-driven pipeline through Inngest, which fetches the PR diff, runs it through an AI review model, and posts inline comments back to GitHub automatically — all within ~30 seconds. Designed a production deployment on AWS EC2 with PM2 process management, Nginx reverse proxy, and Certbot-issued SSL, alongside webhook signature verification and env-based configuration for dev/production separation.

Live in production, automatically reviewing pull requests end-to-end — from webhook receipt to AI-generated inline comments — with zero manual intervention. Currently in beta with GitHub OAuth-based repo connection.