Software architecture · Production AI

I architect production systems where AI is a first-class component — and doesn't break under real load.

For founders and engineering leads shipping AI into products that real customers depend on.

Mobarrez
Software Architect · AI Systems
BasedRemote · Worldwide
AI Systems Architecture
Production Next.js
Agent Infrastructure
System Design

Track record

Six years shipping production software — SaaS platforms, real-time data systems, and the agent infrastructure that runs behind them.

What I do

Three problems worth paying to solve

The demo was flawless. Then it met real users, real costs, and real edge cases.
AI features that survive production
I design agent loops with explicit failure boundaries, bounded retries, and hard cost ceilings. When something upstream misbehaves — and it will — the feature degrades in a way you chose, instead of collapsing in a way you didn't.
- Agent loops and tool orchestration
- Failure boundaries, retries, and fallbacks
- Token, latency, and cost control
It works fine at a hundred users. Nobody can tell you what happens at a hundred thousand.
Architecture that scales
System design, data modelling, and deployment decided deliberately and written down. Growth becomes a capacity question you can answer with a number — not a rewrite you discover halfway through a funding round.
- System and API design
- Data modelling and migration strategy
- Deployment, observability, and rollback
You need the thing built and shipped, not another document explaining how it could be built.
Full product delivery
End-to-end ownership from architecture through to a product in front of users. One person accountable for the decisions, the code, and whether it actually holds up after launch.
- Next.js and TypeScript product engineering
- Backend, infrastructure, and integrations
- Architecture through to launch and handover

Selected work

Three builds, in depth

Problem, approach, outcome — and the trade-off I made to get there. Three is the ceiling on purpose.

Project 012025
TrackTribe
Apex team
A platform to participate in or create sports and social events.
Problem
People in Iran want to find and join local sports events, but existing tools require manual coordination across Telegram groups, separate maps, and handwritten signup lists. Organizers had no single space to post events, manage attendance, and track participant activity.
Approach
Built a role-based platform with three dedicated surfaces — athlete, coach, and organizer — each with its own toolset. Mapbox handles route discovery and event geolocation; a NestJS backend manages user hierarchy, event lifecycle, and activity logging. Postgres (Neon) stores all relational data: events, registrations, tracked sessions, and role permissions.
Outcome
Launched in beta with 50+ events posted, 342+ km of activity tracked across the community, and three fully functional user roles. The platform became the primary event coordination tool for the Apex team's running and cycling community in Tehran.
Trade-off
Supporting three distinct roles from day one meant tripling the permission model and UI surface before proving traction on any single role. A single-role athlete MVP with organizer tools added later would have been faster to ship, but the network effect needed both sides of the marketplace live at once.
Next.js
Mapbox
NestJS
Postgres
Neon
Visit project
Project 022023
Digilogbook
Parbaz
A platform for paraglider pilots to log flights, verify certifications, and connect with coaches — built to replace a manual system where falsified records were common.
Problem
Paraglider certification in Iran relied on paper logs and handwritten attestations, making it easy for unqualified pilots to falsify records and receive certificates they did not earn. Coaches had no verifiable flight history for their students, and the certification body had no audit trail — the opacity protected intermediaries who profited from the broken process.
Approach
Built a digital logbook with verified flight entries, certification tracking, and a coach-student connection layer. Next.js with React Query powered the frontend; a .NET Core backend handled the verification pipeline, certificate eligibility calculations, and the audit trail that would have made the old system's opacity impossible.
Outcome
Delivered over two years as the front-end developer at Parbaz. The software was production-ready, but the company shut down when government support was withdrawn — the system's transparency threatened the intermediaries whose income depended on the old manual process.
Trade-off
The system's viability depended on institutional adoption from a government body that stood to lose from its existence. Building on top of a single approval gate without an independent path to users meant the project died with the relationship — the code was ready, but the route to pilots was gatekept.
Next.js
React Query
.NET Core
Visit project

About

How I think about building software

Most systems don't fail because someone picked the wrong framework. They fail because nobody decided, on purpose, what happens when a dependency is slow, a model returns nonsense, or traffic arrives ten times faster than planned.

I work on those decisions first. That means designing failure boundaries before features, modelling data before writing endpoints, and being explicit about cost — of tokens, of queries, and of the complexity a choice locks you into.

The result is software you can reason about a year later: systems that degrade predictably, scale for understood reasons, and can be handed to a team without an archaeology phase.

For larger builds I work with a small trusted team.

YouTube

I reason about this in public

Architecture decisions, AI systems, and production trade-offs — worked through end to end. The fastest way to judge how I think before we ever talk.

Let's talk about what you're building

Book a call

30 minutes

Thirty minutes, no pitch deck. Bring the problem — the messier the better — and you'll leave with a straight answer on whether I'm the right person for it.

I take on a small number of engagements at a time so each one gets real attention.

Loading calendar…

Prefer to write first?

Send the outline and I'll reply within two business days.

I architect production systems where AI is a first-class component — and doesn't break under real load.

AI features that survive production

Architecture that scales

Full product delivery

TrackTribe

Digilogbook

Book a call

Prefer to write first?