Ahmad Raza

I build real-time dispatch platforms, commerce ERPs and LLM phone agents on MERN and PERN — plus the queues, retries and fallbacks that keep them running when Stripe, Shopify or a model provider doesn't.

See the systemsEmail me

Shipped & running

Next: try to make the phone agent guess

Call replay · Certified Towing phone agent

It can’t quote a balanceit didn’t look up.

Asked a follow-up, a language model answers from memory. So on any turn where the phone agent could state a vehicle fact, the lookup is forced before a word is said. Switch the guard off to hear what it says instead.

Replay · impound-lot callturn 5 of 9
  1. agentCertified Tow.
  2. callerHi — my car got towed last night. Do you have it?
  3. agentI can check. Are you the registered owner?heard at 640 ms
  4. callerYeah, it's mine.
  5. agentWhat's the plate number?heard at 610 ms
  6. callerEight, K, X, M, two, one, four.
  7. agentThat's 8, K, X, M, 2, 1, 4 — is that right?heard at 650 ms
  8. caller

Server · turn 5

caller speaking — waiting for the end of the turn…

Fallback only happens before the first token. After that, half an answer is already being spoken — so a failure there ends the turn instead.

Scripted replay — plate, car and figures invented. Turn logic and timings follow the production agent.

01.

Every integration is a promise that someone else's system stays up. My job is making that not matter.

Connecting systems is the easy half. The real work is what happens when one is down, slow, or sends the same event twice — that's where idempotent jobs, bounded caches, concurrency gating and fallback chains stop being vocabulary and start being the product.

0

offline tests covering the voice agent

0

third-party systems integrated

0

LLM providers behind one fallback chain

0

records per batched cron page

MERN / PERN Stack

Both halves in production: PERN behind FoxTow's dispatch platform, MERN behind Wide ERP.

  • React.js
  • Next.js
  • Node.js
  • Express.js
  • MongoDB
  • PostgreSQL
  • Prisma
  • TypeScript
  • React Query
  • Zustand

Applied AI

LLM features built to degrade, not collapse — fallback chains, streamed responses, and tests that catch silent regressions.

  • Groq → DeepSeek → Gemini fallback
  • OpenAI-compatible SSE
  • Server-side function calling
  • Live voice pipeline
  • Prompt pipelines
  • Offline eval suites

System Design

The decisions you don't get to revisit once real traffic arrives.

  • Multi-tenant schemas
  • Queue workers (RQ)
  • Idempotent jobs
  • Retry & backoff
  • TTL caching
  • Concurrency gating
  • Compound & partial indexes
  • Webhook lifecycles

The 16 systems behind that number— hover to pause

  • StripeStripe
  • ShopifyShopify
  • QuickBooksQuickBooks
  • RingCentral
  • Google MapsGoogle Maps
  • Optimus GPS
  • CloudflareCloudflare R2
  • Google CloudGoogle Cloud Storage
  • Vapi
  • Deepgram
  • Groq
  • DeepSeek
  • Google GeminiGemini
  • PostEx
  • InstaWorld
  • Rocket

02.

MERN Stack Developer @ Webbuggs

April 2025 — Present · Lahore, Pakistan

  • Integrated payments, accounting, telephony, mapping and LLM provider APIs over OAuth and webhooks.
  • Designed multi-tenant PostgreSQL schemas and real-time Socket.io services powering live dispatch operations.
  • Built queue workers, idempotent jobs, retry and backoff, and TTL caching to keep webhook endpoints responsive.
  • Optimized API response times by restructuring database queries and indexing hot read paths.
  • Drove code reviews and turned stakeholder requirements into production-ready releases delivered on schedule.

03.

IN PRODUCTION

AI Call Intake · Live Phone Agent

Certified Towing AI Dispatch

A FastAPI brain behind a Vapi phone agent that answers impound-lot calls: confirms the caller's plate, looks the vehicle up in FoxTow mid-call, and states only what the record says, handing tow requests and anything unclear to a person.

  • Webhooks return immediately and hand off to Redis Queue workers; recording jobs run idempotently with retry and backoff, so a duplicate delivery costs nothing.
  • An OpenAI-compatible SSE endpoint powers the live voice agent with server-side function calling, covered by 480 offline tests.
  • RingCentral over server-to-server JWT with cached tokens and verified webhook lifecycles, behind a Groq → DeepSeek → Gemini chain so a provider outage degrades instead of dropping the line.
  • Python
  • FastAPI
  • PostgreSQL
  • Redis (RQ)
  • Docker
  • Integrations
  • RingCentral
  • Vapi
  • Deepgram
  • Groq
  • DeepSeek
  • Gemini
  • SSE
LIVE

Real-Time Dispatch Platform

FoxTow

Multi-tenant towing dispatch: live truck telematics, driver assignment and an interactive map that has to stay correct while trucks move.

  • Socket.io dispatch loop — driver registration, call assignment, accept/reject, duty-shift changes and live location events streaming to the dispatcher map.
  • Stripe subscription billing with webhooks, QuickBooks Online over OAuth, RingCentral SMS and voice, Google Geocoding, and Optimus GPS telematics — six systems, one tenancy model.
  • Reliability work across the multi-tenant schema: bounded geocode caching with TTL, concurrency gating with retry/backoff, batched cron jobs paginating 500–1000 records, and compound and partial indexes on hot read paths.
  • PostgreSQL
  • Prisma
  • Node.js
  • Express.js
  • React.js
  • TypeScript
  • Vite
  • Socket.io
  • Integrations
  • Stripe
  • QuickBooks
  • RingCentral
  • Google Maps
  • Optimus GPS
  • Cloudflare R2
LIVE

Commerce & Inventory

Wide ERP POS

An ERP and point-of-sale system holding a Shopify catalogue and three courier networks in sync behind a single dashboard.

  • Shopify over OAuth and the Admin GraphQL and REST APIs for order, product and fulfillment sync, plus PostEx, InstaWorld and Rocket for city lookup, shipment creation and tracking.
  • A resumable background sync worker that cursor-paginates Shopify products, pipes images through Google Cloud Storage, and bulk-inserts with unordered writes and capped retries — safe to interrupt, safe to resume.
  • Server and client state kept separate — React Query for cached server state, Zustand for client state — over MongoDB aggregation pipelines driving sales, inventory and staff dashboards.
  • Next.js
  • MongoDB
  • Express.js
  • Node.js
  • TypeScript
  • ShadCN UI
  • React Query
  • Zustand
  • Integrations
  • Shopify Admin API
  • PostEx
  • InstaWorld
  • Rocket
  • Google Cloud Storage

04.

I'm open to interesting problems — contract work, full-time roles, or just comparing notes on shipping AI features that survive real traffic.

Say hello