Muhammad RashidPlatform & AI Engineer

Ibuildsystemsthatstayup.AndtheAIthatrunsinsidethem.

Technical lead for a live commerce and logistics platform in Mexico. Terraform-managed AWS, keyless CI/CD, vision models on GPUs with vLLM. And 86 videos teaching 37.2K engineers how it's done.

See the systemsWork with meWatch on YouTube ↗
status · October 2026
  • $ status --watch
  • logistika-urbanaecs fargate · 4 services · rds multi-azup
  • skin-analysisclaude | openai · json schema · per-store capup
  • kyc-vlmvllm · qwen3-vl-8b + paddleocr-vl · 1× L40Swarm
  • private-llm-chatopen source · MIT · sagemaker scale-to-zeroup
  • codewithmuh37.2K subscribers · 86 videos · 82 postslive
  • $
00A note before the work

Most portfolios show demos. This one shows systems in production: a commerce and logistics platform in Mexico, vision models on GPUs, an open-source stack anyone can run. I design the infrastructure, write the services, ship the apps, and run the models in between. Then I teach it. 

01Systems in production

Five systems. Real stores, real GPUs, real traffic.

Private client systems are described, not linked. The open-source ones link straight to the code and the running thing.

01/ 05Commerce + last-mile logistics · Mexico

Logistika Urbana

Technical Lead (contract) · 2024 — present

A multi-brand retail, wholesale and delivery platform: Yoon Lab K-beauty stores online and on the street, Fleksi delivery, a driver network. Built and run end to end by a two-person team.

  • 26 domain modules, ~535 API routes, 436 automated tests on the core backend
  • 11 Terraform modules: VPC, ALB + WAFv2, ECS Fargate, RDS Multi-AZ, Redis, OpenSearch, GuardDuty
  • EC2 → Fargate migration (March 2026) with deployment circuit breakers and automatic rollback
  • Keyless OIDC CI/CD: SHA-tagged images, migrations as one-off tasks, a TLS-validated health gate
  • Native driver apps in Kotlin and SwiftUI, an offline-first POS, kiosks, a React admin
Django / DRFCeleryPostgreSQLRedisOpenSearchTerraformECS FargateStripeKotlinSwiftUI
fig. 01production topology · aws mexico
CLIENTSStorefrontwebAdminreact · amplifyDriver appkotlin · swiftuiPOSoffline-firstKioskin-storeALB + WAFv2tls · managed rulesterraform · 11 modulesoidc deploys · auto-rollbackECS FARGATE · AUTOSCALE 2–6API (ASGI)django · channelsCelery worker47 tasksCelery beat22 schedulesOSRMself-hosted routingcircuit breaker · rolling deploysEXTERNALStripeCFDIEnviaTwilioTikTok ShopAmazonDATA · AWS MEXICORDS Postgresmulti-azRediscache · brokerOpenSearchcatalogS3signed urls
~2,300 commits19 repositories4 Fargate services
02/ 05Vision LLM · Yoon Lab

AI skin analysis, in a retail store

Production AI feature · 2026

Customers scan an in-store QR code, take a photo, and get a personalised routine built only from products that are actually in stock. Claude and OpenAI sit behind a provider layer that can be switched at runtime.

  • JSON-schema structured outputs, one-hour prompt caching, refusal fallbacks
  • An evaluation harness sweeps model × reasoning effort against independently labelled photos, with pass gates written before the results
  • Per-device, per-email, per-/24 and per-store quotas; HMAC-signed QR tokens; a per-store daily cap is the hard ceiling on spend
  • No facial images stored, by design (Mexican LFPDPPP)
Claude APIOpenAI ResponsesJSON SchemaPrompt cachingDjangoHMACEvals
fig. 02request path · controls
In-store QRhmac tokenPhotonever storedPROVIDER LAYER · RUNTIME SWITCHClaudeactiveOpenAIstandby1-hour prompt caching · refusal fallbacksJSON schemastructured outputRoutinepersonalisedLive catalogin stock onlyEVALUATION HARNESSmodel × reasoning effortagreement vs inter-rater ceilingcost per scan · unusable ratepass gates written before resultsSPEND & ABUSE CONTROLSdeviceemail/24storeper-store daily cap = hard ceiling on spend
runtime-switchable providerhard spend ceilingzero image retention
03/ 05Open-weight vision models on one GPU

Zero-retention AI KYC

Fintech client · 2026

Identity documents go in, structured fields come out, and nothing is kept. Two vision-language models share a single L40S behind a custom SageMaker router.

  • Qwen3-VL-8B (FP8) and PaddleOCR-VL served with vLLM on one 48 GB L40S, roughly a 75/25 memory split
  • A FastAPI router implementing SageMaker's /ping and /invocations contract; weights loaded offline from S3
  • 12 Terraform modules with content-addressed endpoint configs, so a no-op apply never triggers a 15-minute reload
  • A GPU lifecycle controller: always-on, scheduled, on-demand with idle stop, or off. Never mid-document
  • Data Capture disabled and asserted by test; PII-redaction invariants gate CI
vLLMSageMakerQwen3-VLPaddleOCR-VLFastAPITerraformECRGitHub Actions OIDC
Read the write-up →
fig. 03inference path · gpu lifecycle
ID documentimage · pdfFastAPI router/ping · /invocationsterraform · 12 modulescontent-addressed endpoint configsS3 weightsoffline loadSAGEMAKER · ML.G6E.XLARGE · 1× L40S 48 GBQwen3-VL-8B · FP8vLLM · fields & reasoningPaddleOCR-VLvLLM · OCR~75% gpu memory~25%Fieldsjson · validatednothing retainedGPU LIFECYCLE CONTROLLERalways-onscheduledon-demandoffidle stop after 30 min · never mid-documentCOMPLIANCE GATESdata capture: off (asserted by test)pii-redaction invariants gate cinetwork-isolated endpoint
1 GPU, 2 models12 Terraform modules~10 GB inference image
04/ 05AI policy & compliance analysis · US SaaS

SYBAL

Lead Backend Engineer, Platform & ML Integration · 2025 — 2026

Took a data-science team's GPU models and made them a product: async inference, durable result ingestion, and a database per tenant.

  • Qwen2.5-3B LoRA rule extractor and RoBERTa NER on SageMaker async endpoints, autoscaling 0 → 5
  • Continuation-token batching and partial-result merging for documents too large for one invocation
  • Celery ingestion with auto-save, 40 retries and a five-minute sweeper that re-queues anything pending
  • Database-per-tenant on Aurora PostgreSQL: a Django DB router, async provisioning, RDS Proxy
  • Top contributor: 338 of 1,109 commits in six months
DjangoCelerySageMaker asyncAurora PostgreSQLRDS ProxyFAISSDockerGitHub Actions
fig. 04ml integration · tenancy
TENANTStenant atenant btenant cDjango APIdb router · celeryS3 staginginputs · outputsSAGEMAKER ASYNC · AUTOSCALE 0 → 5Qwen2.5-3B LoRArule extractorRoBERTa NERentity extractioncontinuation tokens · partial mergeCelery ingestionauto-save · 40 retries · sweeper5-minute sweeper re-queues pendingper-tenant anomaly detectionAURORA POSTGRESQLrds proxyddl endpointdb_adb_bdb_cdatabase per tenant · encrypted credentials
db per tenantasync GPU inferencedurable ingestion
05/ 05Open source · MIT

Private LLM Chat

Creator & maintainer · 2026 — present

A self-hosted ChatGPT-style platform for open-weight models on your own GPU. Several models per GPU, weights from S3, an optional no-internet VPC, and a GPU that switches itself off when idle.

  • Qwen, Llama, Gemma and DeepSeek served with vLLM on Amazon SageMaker, scaling to zero
  • Django API and React app: streaming chat, documents and images with OCR, Whisper voice input, web search, Google sign-in with 2FA
  • Terraform for the whole stack: ALB, CDN, ECS, SageMaker, database, network, ECR, CI/CD
  • A laptop-first path with Docker Model Runner; a Google Cloud Run (GPU) port in progress
  • A docs site with guides, architecture, GPU sizing, cost and API reference
vLLMSageMakerTerraformDjangoReactWhisperS3Cloud Run
llm.codewithmuh.com ↗Source on GitHub ↗
fig. 05self-hosted stack
REACT APPstream · docs · images · voiceDjango APIstreaming rest · 2faWhisperOCRWeb searchSAGEMAKER · VLLMQwenLlamaGemmaDeepSeekGPUseveral models per gpuscale → 0 when idleweights from s3 · no-internet vpc optionS3 weightsofflinePrivate VPCoptionalModel Runnerlaptop · dockerTERRAFORM · ONE STACKalbcdnecssagemakerdatabasenetworkecremailcicddocs site · guides · gpu sizing · cost reference · MIT
MITscale to zerodocs site
02Counted, not rounded up

as of October 2026 · github · youtube · upwork

Repositories0128 public · github.com/codewithmuh
Forks0420 stars across public repos
Subscribers0.0KYouTube · @codewithmuh
Videos0AI engineering, AWS, Terraform, LLM apps
Articles02022 → 2026 · codewithmuh.com
AWS certifications0Developer Associate · AI · Cloud Practitioner
Contracts0+Upwork · Top Rated · 100% job success
Years in production0+Python, AWS, mobile, models
Featured byHostingerReplitConsensusNexos.aiAWS Community Builder
03One person, the whole stack

From the screen in the store to the GPU in the region.

Four layers, one owner. That is what lets a two-person team run a platform that would normally take a department.

01Surfaces· What people touch
01

Surfaces

Storefronts, admin, driver apps, store operations, kiosks.

  • Next.js / React
  • TypeScript
  • Kotlin · Jetpack Compose
  • Swift · SwiftUI
  • Offline-first POS
  • React Native
02

Services

The largest one: 26 modules, ~535 routes, 436 tests.

  • Python
  • Django / DRF
  • Channels (WebSockets)
  • Celery + beat
  • FastAPI
  • PostgreSQL
  • Redis
  • OpenSearch
  • Stripe
  • Twilio
03

Models

Structured outputs, prompt caching, evals, quotas. I run models; I don't train them.

  • Claude API
  • OpenAI
  • vLLM
  • SageMaker endpoints
  • Qwen3-VL · PaddleOCR-VL
  • pgvector RAG
  • LangGraph
  • Vapi · ElevenLabs
04

Platform

EKS, Helm and Argo CD live in the open-source reference builds.

  • Terraform
  • AWS ECS Fargate
  • RDS Multi-AZ
  • ALB · WAFv2
  • Secrets Manager
  • GuardDuty
  • GitHub Actions (OIDC)
  • CloudWatch
  • Prometheus · Grafana · Loki
  • gitleaks
04Seven years, in order

Freelancer, founder, CTO, technical lead.

Every step was remote, from Lahore, for teams in Portugal, the United States and Mexico.

  1. 2019

    Fiverr: 100+ machine-learning and data projects while at university

    Level 2 Seller

  2. 2020

    M.Sc. Computer Science, University of the Punjab

    Lahore

  3. 2020

    Founded LozpData; started on Upwork

    Top Rated · 100% Job Success

  4. 2022

    Started writing. The first of 82 posts

    codewithmuh.com

  5. 2023

    APOLOP2P: backend engineer, promoted to CTO

    Portugal · remote

  6. 2023

    CodeWithMuh on YouTube

    37.2K subscribers today

  7. 2024

    Logistika Urbana: technical lead for a live commerce and logistics platform

    Mexico · ongoing

  8. 2024

    Taught a two-month DevOps program to 15+ US students

    Terraform · Docker · Kubernetes · AWS

  9. 2025

    Three AWS certifications; AWS Community Builder

    Developer Associate · AI · Cloud Practitioner

  10. 2025

    SYBAL: lead backend engineer, platform & ML integration

    United States · 2025 — 2026

  11. 2026

    Founded AISeekho.pk; shipped AI KYC on vLLM; open-sourced Private LLM Chat

    the year of running models

  12. 2027

    Tokyo

    relocating

05Teaching

I teach what I run.

37.2K subscribers · 86 videos · 82 articles · a 1,700-reader newsletter. Everything on the channel comes from a system that is actually live.

06Open source & products

Built in public. Run it yourself.

128 public repositories, 420 stars, 611 forks. Every product ships with a live URL and a repository you can clone.

three-tier-devsecops-project★ 22⑂ 134Terraform · K8sai-voice-agent★ 22⑂ 10Pythonreact-aws-ec2-nginx★ 19⑂ 41JavaScriptdjango-starter-kit★ 18⑂ 9Pythonblogprojectdrf★ 18⑂ 93Pythonai-support-agent★ 14⑂ 5Pythonci-cd-configurations-files★ 13⑂ 26YAMLbackend_course★ 12⑂ 28Pythonhearthline★ 8⑂ 5TypeScript · Pythonreact-aws-eks-github-actions★ 8⑂ 25Terraformdjango-aws-ec2-autoscaling★ 7⑂ 15Pythonprivate-llm-chat-sagemaker-aws★ 1⑂ 0Terraform · Python
  • 01

    Cartlift

    Ecommerce growth daemon: a four-lens audit in ~30 s, drafted variants shipped through a 3 KB snippet, a Thompson-sampling allocator that pins the winner.

    Live · MITLive ↗Code ↗More →
  • 02

    Hearthline

    The 24/7 AI front desk for home-service businesses: answers calls, qualifies, books, drafts photo-based quotes. Ten UI languages, Japanese included.

    BetaLive ↗Code ↗More →
  • 03

    Private LLM Chat

    Self-hosted chat for open-weight models on your own GPU. vLLM on SageMaker, Terraform for everything, scale to zero.

    Open source · MITLive ↗Code ↗
  • 04

    AI Receptionist

    Voice agent on Vapi with a Claude tool-use loop: calendar booking, Twilio SMS, call logging.

    Open source · MITCode ↗More →
  • 05

    AI Support Agent

    WhatsApp, Telegram, Gmail and web support with Haiku triage, Sonnet answers, pgvector retrieval and human escalation.

    Open sourceCode ↗
  • 06

    AI Hiring Agent

    Claude resume scoring, a voice screening call, transcript analysis. Django and Next.js.

    Open sourceCode ↗
  • 07

    Three-Tier DevSecOps on EKS

    The reference pipeline: Terraform, Jenkins, Argo CD GitOps, Trivy and SonarQube scanning, Prometheus and Grafana.

    Open source · MITCode ↗More →
Startups →AI agents →DevOps projects →All 128 public repos ↗
07Next

Let's build something that stays up.

contact@codewithmuh.com
Book a build sessionWhatsAppMedia kit
Lahore, Pakistan · ––:––Remote, worldwide · senior platform, AI and lead rolesRelocating to Tokyo · 2027