Rehan Bashir

Full-Stack Software Engineer / AI Systems

I build production software where AI, realtime systems, and dependable backend engineering meet.

Two years of shipping full-stack products — most recently a multi-tenant healthcare platform taken from first commit to production. I work across the interface, the API, the data layer, and the behaviour of the model in the loop, and I care most about the parts that have to hold up under real traffic.

  • Healthcare SaaS
  • Voice AI
  • RAG
  • Realtime systems
  • Multi-tenant architecture

Virginia, USA

The shape of the work

Across the stack,not inside one layer.

I tend to work across the stack rather than inside one layer of it — interface, API, database, and the model in the loop. The work I'm proudest of lives where those layers meet a production constraint: a latency budget, a permission boundary, a test that has to prove isolation actually holds.

B.S. Software Engineering — FAST National University, 2020–2024

Experience

Oneproduction system after another.

Jun 2025 – Apr 2026 · New York, NY

Full-Stack Software Engineer

Software Pro

I joined early and helped take a multi-tenant healthcare SaaS from first commits to a system clinics depend on. Most of my time went to two hard edges: syncing patient records across Epic, Cerner, and ModMed without letting tenants bleed into each other, and standing up a voice-agent layer on Retell AI that answers in under a second using retrieval through Pinecone. The compliance work was part of shipping, not a phase bolted on afterward.

  • EMR sync across three vendor systems — Epic, Cerner, ModMed
  • Sub-1s voice agent on Retell AI with Pinecone retrieval
  • HIPAA hardening: PHI redaction, OAuth 2.0, role-based access, audit logging

Node.js · React · TypeScript · MongoDB Atlas · Pinecone · Redis · Twilio

Role 1

Dec 2024 – Jun 2025

Contract Software Engineer

University Research Lab

I built research-facing web apps — collection and analysis tools on React, TypeScript, Vite, and Firebase — for teams working with large text datasets. That included the NLP pipelines and LLM-backed features they needed to make sense of the corpora as they grew.

  • Collection and analysis apps for large text datasets
  • NLP pipelines and LLM-backed analysis features

React · TypeScript · Vite · Firebase

Role 2

Jul 2024 – Sep 2024

Backend Engineer Intern

PakWheels.com

I owned backend APIs for a car-inspection booking system running more than 500 appointments a day. Alongside feature work I cleared over 20 critical issues, and a focused refactor of the legacy backend brought a 15% performance gain.

  • REST APIs sustaining 500+ daily bookings
  • 20+ critical backend issues resolved
  • 15% performance gain from a legacy refactor

Role 3

Aug 2023 – Aug 2024 · Remote

AI Engineer

UrduX

I built DAOChatbot end to end — a retrieval-augmented query system on Flask, React, MongoDB, and Pinecone — and owned Gender-DEI Bot, a DistilBERT bias-detection platform. On the latter I cut inference latency by 60%.

  • End-to-end RAG system, DAOChatbot
  • DistilBERT bias-detection platform, owned
  • 60% inference-latency reduction

Flask · React · MongoDB · Pinecone · DistilBERT

Role 4

The Developer Guys

Founder chapter

The Developer Guys

Co-Founder & Lead Engineer · Jun 2026 – Present

An AI and full-stack practice I started. The focus is production LLM integration and healthcare-adjacent workflows, and the work that comes up most in client conversations is reference implementations — RAG systems, voice agents, and secure multi-tenant architecture.

  • Web Apps
  • APIs
  • RAG
  • Voice
  • Multi-tenant Systems

Studio work

  • The Developer Guys studio work — Gallus Medix
  • The Developer Guys studio work — UForge
  • The Developer Guys studio work — Gondal Protein Farms

TypeScript · Python · Next.js · PostgreSQL · OpenAI · Anthropic

Selected work

Three systems,built end to end.

Personal projects where the interesting part is buried in the architecture — telemetry timing, database-enforced permissions, inventory that cannot oversell.

Formula 1 car cockpit, close up

01Personal project

F1 25 AI Race Engineer

A voice race engineer for EA's F1 25 that reads the live telemetry feed and only speaks when the race actually changes.

  • 16UDP packet types decoded
  • 20+ Hztelemetry decode rate
  • 451regression assertions

What it is

F1 25 streams the car's state over UDP. The engineer decodes all 16 packet types at 20+ Hz, keeps 11 stateful assessments of how the race is going, and talks to the driver over the OpenAI Realtime API — but only on a real state change, never on a timer.

What made it hard

A language model will happily invent a lap time. Every number the driver hears is fetched through a tool call first, so the model reports measured telemetry instead of guessing it. A 451-assertion harness replays the whole pipeline from synthesized packets, so it runs with no game open.

16 UDP packet types

0 / 16 decoded

Race state · 20+ Hz

Speaks on a real state change — never on a timer.

  • All 16 UDP packet types decoded at 20+ Hz
  • 11-assessment situation engine decides when it speaks
  • Driver-facing replies routed through tool calls — no invented numbers
  • 451-assertion regression harness runs headless from synthesized packets
PythonasyncioFastAPIOpenAI Realtime API

Photo: Adrian Kusznirewicz / Unsplash

Medical professionals working at a laptop

02Personal project

ClinicOS

Clinic management for small multi-specialty practices in Pakistan, where the database itself enforces who can see what.

  • RLSon every query
  • 57real-Postgres integration tests
  • 3isolated roles

What it is

Reception books and checks in patients, doctors write notes, billing runs invoices — and none of them can see each other's data. Row-Level Security gates every query at the database, not the interface. Consultations can be dictated in mixed Urdu and English and come back as an editable draft.

What made it hard

Proving the isolation actually holds. The integration suite runs against a real Postgres instance and tries to break through — spoofing tenant and role fields, reading across roles — and confirms the database rejects each attempt.

Row-Level Security · resolved per query

Reception

  • Scheduling···
  • Clinical notes···
  • Billing···

Doctor

  • Scheduling···
  • Clinical notes···
  • Billing···

Billing

  • Scheduling···
  • Clinical notes···
  • Billing···

Reception can’t read notes. Doctors can’t see billing. The database refuses the row — not the screen.

  • Row-Level Security enforces access on every read and write
  • Reception can't open clinical notes; doctors can't open billing
  • 57 real-Postgres tests for cross-role isolation and spoofed-field rejection
  • Code-switched Urdu/English dictation becomes an editable consultation draft
Next.js 16React 19TypeScriptSupabasePostgreSQL

Photo: Nadezhda Moryak / Pexels

A group of supplement bottles

03Personal project

VitaCommerce

A supplements commerce monorepo built around inventory that can't oversell — even when two checkouts race for the last unit.

  • 14API endpoints
  • 4surfaces, one monorepo
  • 7-stateorder machine

What it is

Storefront, admin, a 14-endpoint API, and an Expo app in one repo. Stock is tracked per lot and allocated first-expired-first-out, every movement is written to an append-only ledger, and orders move through a seven-state machine.

What made it hard

Concurrency. Inventory commits in two phases and quantity guards hold the line when checkouts collide, so the last unit sells exactly once.

FEFO allocation · earliest expiry leaves first

7-state order machine

Two checkouts race for the last unit

  • Per-lot first-expired-first-out allocation with an append-only movement ledger
  • Two-phase inventory: reserve, then commit
  • Quantity guards block oversell under racing checkouts
  • Storefront, admin, 14-endpoint API, and Expo app in one monorepo
TypeScriptNext.js 16PrismaPostgreSQLExpo

Photo: Natali Hordiiuk / Unsplash