Senior Full Stack Engineer and AI Architect. I build LLM systems and multi-tenant platforms that hold up in production.
Eight years and 500+ projects delivered across fintech, SaaS, e-commerce and ed-tech, with teams in the USA, Europe and the Middle East. Python, TypeScript, and AI pipelines that still behave on the ten-thousandth call.
Worked withNuvitaFixFinanzBullseyes.aiOverZakiFlySmartDealsG3MSPoshTextilesDigital Media Flow
Why hire me
The problem & the answer
Anyone can generate code now.The expensive part is knowing why it broke.
I spent eight years debugging before AI existed. That is what you are actually hiring.
2018–2022No model to askYears of reading stack traces, reasoning about state, and finding the real cause instead of the first plausible one. There was nothing to ask.
2023–2024AI enters the workflowGenerated code arrives fast and looks right. I already knew what wrong looked like, so it never went in unreviewed.
2025–2026Building AI systems, not just using themLLM vision pipelines, safety filters, deterministic fallbacks. Same engineering standard, applied to a new kind of system.
Two reasons, and they are the whole pitch
01
I debug from first principles, not from autocomplete
I have been writing production code since 2018. For most of those years there was nothing to ask, so you learned to trace a bug to its actual cause and fix that. That habit is why I can read generated code and tell you what it got wrong.
A model will hand you a confident, working-looking answer to a problem it has misunderstood. I use AI every day for speed and review its output the way I would review a junior engineer's pull request.
ProofThe Nuvita fallback system and the FixFinanz webhook validation both exist because I did not trust the first version that ran.
02
The language is not the skill. The system is.
Frameworks are syntax on top of the same few ideas: state, data flow, boundaries, and what happens when something fails. Once you understand those, a new language is a week of reading, not a career change.
That is why I can join a team on any stack and be useful in days. I do not need the job to match the tools I already know, because the part I am actually good at does not live in the tools.
ProofReact frontend in 2018, Firebase and Node backends by 2023, NestJS and TypeScript through 2025, Python and FastAPI in 2026. Same standard every time.
Appreciation letter received from the President of Pakistan
B.S. Computer Science Engineering
GC University, Pakistan · 2015–2019
Zero privilege-escalation incidents
Across a year of production RBAC work at FixFinanz, with 90%+ test coverage held throughout
Three problems founders hire me for
What I do
01
AI & LLM product engineering
Vision pipelines, structured output, retrieval, coaching engines. The demo is the easy part. I build the guardrails that keep it correct once real users hit it.
Strict JSON-schema structured output
External cross-validation of model answers
Deterministic fallback on model failure
TTL caching and per-user usage limits
02
Full-stack product build
From an empty repository to a product with paying customers. Next.js on the front, NestJS or FastAPI behind it, with the boring parts done properly.
Multi-tenant data isolation and RBAC
Stripe billing with idempotent webhooks
Real-time updates over Socket.io
Jest unit and end-to-end coverage
03
Architecture and rescue
For codebases that became slow to change. I draw the service boundaries, add the tests that make refactors safe, and hand back something your next engineer can read.
Modular monolith boundaries
State normalization and render profiling
Webhook integrity and replay safety
Documentation your team keeps using
AI agents and automation
Built to survive contact with users
A demo takes an afternoon. A system that runs ten thousand times against real people, with real money and real consequences attached, is a different product. Everything below is something I have shipped in production, with the project it shipped in named next to it.
Schema-enforced output
A strict JSON schema on every model response, validated before anything downstream is allowed to touch it. No parsing free text and hoping.
Nuvita · Pydantic v2 strict validation on every LLM and database interaction
External validation of model output
The model detects, an authoritative source decides. Anything the model asserts that can be checked against real data gets checked before a user sees it.
Nuvita · every detected food cross-validated against USDA FoodData Central over async HTTPX
Deterministic fallback
When the model fails, times out or returns nonsense, the feature degrades to a rule engine instead of disappearing. Users never see a dead screen.
Filters in front of generated text for domains where the wrong wording is a liability rather than an inconvenience.
Nuvita · medical-language safety filter on all coaching output
Event pipelines and ingestion
Webhook pipelines that ingest, verify, classify and route automatically, with signature checks and idempotency so replays never corrupt your data.
FixFinanz · Meta Graph API ingestion, HMAC SHA-256 validation, 1,000+ leads/month, zero integrity failures
Vision and document extraction
Turning a photo, a scan or a document into structured, validated data that the rest of your system can rely on.
Nuvita · LLM Vision pipeline plus barcode scanning, zero manual input
Workflow automation in n8n and in code
Automations that connect the tools a team already uses. Built visually in n8n where the client wants to own it afterwards, or in code where it needs to do more than a node can.
Direct client engagements · handed over with a walkthrough so the team can maintain it
AI added to software that already exists
Putting LLM features into a product that already has users and a codebase, without a rewrite and without breaking what already works.
Direct client engagements · integration work across existing stacks
Conversational systems
Chat interfaces wired into real business data and real rules, rather than a model answering from nothing.
Django chatbot delivered on Upwork · five stars, reviewed below
Some direct client work is covered by NDA and is not listed by name here. I can walk through the architecture and the decisions on a call.
Photo-based nutrition tracking already existed, and nobody trusted it. The models guess portion sizes and invent macros, so the numbers look plausible and are wrong. Anything built on top of that is a toy.
The decision
Use the model for detection only, never for the numbers. Live meal photos go through an LLM Vision model with strict JSON-schema structured output, then every detected food is cross-validated against USDA FoodData Central over async HTTPX. Packaged food skips the model entirely through a camera-first barcode scanner reading straight to OpenFoodFacts.
What broke
The first coaching engine called the LLM on every dashboard load, and the cost curve went vertical. I split it in two: a deterministic rule engine handles the actual reasoning (calorie balance, protein deficits, meal timing, streaks) and the model only personalises the wording. A 90-second TTL cache absorbs repeat loads, a medical-language safety filter sits in front of the output, and a transparent fallback covers model failure so the feature degrades instead of disappearing.
The result
A production LLM pipeline running across scanning, coaching, analytics and billing. Stripe freemium with a billing_events audit table and idempotent webhooks, per-user LLM usage limits, Fitbit OAuth with AES-encrypted token storage, and Supabase row-level security isolating every user's data.
Advisors were losing leads in the gaps between Meta ad forms, spreadsheets and personal calendars. Three roles — customer, advisor, admin — needed one platform with very different permissions, and the compliance bar in German financial advisory leaves no room for a leaky one.
The decision
A NestJS modular monolith — auth, appointments, pipelines, contracts — with strict service boundaries, instead of microservices a team that size would spend a year operating. Role-scoped dashboards in Next.js with route-level guards and permission-aligned UI, so the interface never renders an action the API would reject.
What broke
Meta's Graph API replays webhooks, and the first ingestion run created duplicate leads from the same submission. I moved validation to the boundary: HMAC SHA-256 signature checks and DTO mapping before anything touches a service, with idempotent ingestion behind it. Booking was the second one — two advisors could take the same slot, so the flow became an explicit state machine with slot-policy and calendar enforcement.
The result
Cross-module coupling down roughly 60%. Over 1,000 leads a month ingested with zero data-integrity failures. Scheduling conflicts down about 80%. High-traffic screens about 45% faster after Redux Toolkit entity normalization and targeted cache updates. 90%+ test coverage held for the length of the engagement, and no privilege-escalation incidents in production.
A logistics platform along the lines of Microsoft Dynamics Business Central, running warehouse operations and multi-carrier shipping for a US textile exporter.
JavaScriptReactNode.jsBusiness Central APIFedEx APIUPS APISTAMPAuthorize.Net
The problem
Shipments were created in email, picked from printed sheets, and labelled by typing addresses into three separate carrier portals. Every typo became a returned pallet and an argument about who paid for it.
The decision
One system covering shipment creation, checking, picking, packing and dispatch, tied to Microsoft Dynamics Business Central through the Microsoft APIs so every incoming order is matched against the ERP record before approval. FedEx, UPS and STAMP integrated directly rather than through a reseller, so address validation and label printing happen inside the same flow as the pick.
What broke
The carriers disagree about addresses. The same street would validate at UPS and fail at FedEx, and we were finding out after buying the label. Validation moved to the point of entry, before a label is ever purchased, with the carrier's own normalised address written back to the order.
The result
500+ shipments a month running through the system, and manual shipping errors down 90%. Lot and bin scanning at inspection, inventory across 750 pages of SKUs, bulk CSV import with per-row ERP matching, and Authorize.Net for inter-company settlement. JavaScript is about 99.9% of the codebase.
Digital Media Flow2020–21Full Stack Developer · New YorkNYC agency. Shipped for 10+ clients and moved from frontend specialist to owning whole deliverables within 6 months. Docker and Heroku pipelines at 99.9% uptime.React · Node.js · Docker
X Logic Solutions2018–19Senior React DeveloperLed React work across 10+ client engagements and cut incident response time 40% through structured debugging and code-review practice.React · Code review · Debugging
Plus 500+ freelance and contract projects delivered on and off platform since 2018
People who have worked with me
References
Martin Zenner
CEO, Digitale-Neukunden / FixFinanz · Germany
Reference available on request
Robert Romulus
Co-founder & CEO, Digital Media Flow LLC · New York
I had the pleasure of working with Muhammad and was consistently impressed by both his technical ability and his professionalism. He is a highly skilled Python developer with a strong understanding of software architecture, backend development, APIs, automation, and problem-solving. He just get things very fast.
What sets Muhammed apart is his ability to quickly understand complex business requirements and turn them into clean, scalable, and maintainable solutions. He writes well-structured code, pays close attention to performance and reliability, and always approaches challenges with a thoughtful, solution-oriented mindset.
Beyond his technical expertise, Muhammad is an excellent collaborator. He communicates clearly, works well with cross-functional teams, welcomes feedback, and is always willing to help others. His dedication to delivering high-quality work and his commitment to continuous learning make him a valuable asset to any engineering team.
I would confidently recommend Muhammed to any organization looking for a talented Python developer and Javascript software engineer. He brings technical excellence, integrity, and a positive attitude to every project, and I look forward to seeing his continued success.
★★★★★Every review below is five starsEvery review reproduced below is five stars. Verified on Fiverr and Upwork, from clients in France, Germany, Sweden, the United States, Georgia and Uzbekistan.
Full Stack Developer for Django Chatbot Development★★★★★
Working with Muhammad on this Django chatbot project has been an excellent experience. He is incredibly honest, hardworking, and dedicated to delivering top-quality results.
Committed to qualityDetail orientedProfessional
UpVerified Upwork clientUpwork
★★★★★
Great work and fast. Exactly what I was expecting!
bxav42France · Fiverr
Ongoing collaboration★★★★★
He is a great seller and does his work as best as he can. Recommend him for coding or anything.
pardhunamburiUnited States · Fiverr
Ongoing collaboration★★★★★
Very good works — very fast and bug-free.
martinzennerGermany · Fiverr
Fixed Bugs in React, Redux and React Testing Application★★★★★
Muhammad did an exceptional job resolving complex issues in our React, Redux, and testing environment. He quickly identified and fixed component re-rendering problems, optimized state management, and ensured our unit tests passed.
Solution orientedCommitted to quality
UpVerified Upwork clientUpwork
Ongoing collaboration★★★★★
Very easy to work with. Very good communication.
rileyarunaku609United States · Fiverr
Skilled Next.js and React.js Developer for Advanced Web Application★★★★★
Muhammad’s mastery of React.js for front-end development and seamless API integration were simply outstanding. With impeccable professionalism, he delivered a top-tier product promptly. Highly recommended for his expertise and reliability.
Committed to qualityProfessionalReliable
UpVerified Upwork clientUpwork
Ongoing collaboration★★★★★
It went perfect. Muhammad delivered very fast and I got a perfect product. Will work again with this seller!
victoria_kristiSweden · Fiverr
Ongoing collaboration★★★★★
Amazing! Hamad is a really full stack developer! He shines at both front and back end. Can’t wait to work with him again. From coding to communication, everything was way above par — his knowledge of JS frameworks (React.js and Node.js) is impressive.
allhassan991Georgia · Fiverr
Experienced Next.js 14 Developer for High-Complexity Design Project★★★★★
It was so great to work with Muhammad. Very professional and great quality work. Looking forward to working more.
CollaborativeProfessionalCommitted to quality
UpVerified Upwork clientUpwork
Ongoing collaboration★★★★★
Thank you for your exceptional work as our backend developer using Node.js. Your coding skills and professionalism have been invaluable in delivering high-quality projects.
majlydlyUzbekistan · Fiverr
★★★★★
Best of all, I will keep working with him for life. Great guy with great resources, and a good teammate and colleague.
reignnySweden · Fiverr
How working together goes
Process
01
Intro call
Thirty minutes. You describe the problem, I tell you honestly whether I am the right person for it.
Within 24 hours
02
Architecture note
A short written plan: the approach, what I would not build, the risks, and a timeline you can hold me to.
2–3 days
03
Weekly slices
Something running every week, deployed where you can click it. No month-long silences ending in a surprise.
Ongoing
04
Handover
Tests, documentation and a walkthrough, so the engineer after me can work without calling me.
At close
Get in touch
Tell me what you are building.
Full-time senior and founding engineer roles, or a project with a real deadline. Remote, worldwide. I reply to everything within a day.
