Senior Software Architect · Senior AI Engineer
AI made software fast. I build it to last.
Anyone can ship an AI demo in a weekend now. The hard part comes after: security, data, money, and the six months after launch. That’s the part I architect. Three products I designed and built are running in production today.
How I build
AI didn’t change what software needs. It hid it behind speed.
A tool that works for two weeks and then cracks, or quietly leaves the door open, costs more than never building it. So I build the other way.
- 01
Software still needs a parent.
AI writes code fast, but code without an owner rots. I stay responsible for what I ship, long after the demo got its applause.
In practice: All three products below are live in production. - 02
Armor before speed.
A lot of AI tools shipped today are tanks without armor. Security and data integrity go into the foundation, not into a sprint after the breach.
In practice: WorkPlus members can't touch their own credit balance, even through the API. - 03
The model is a component, not the product.
LLMs get defined tools, limits and retries, inside ordinary engineering. The model decides; plain, testable code does the work.
In practice: In Reelio, the model plans the edit and a deterministic pipeline renders it. - 04
The business first, then the tech.
I start from where it hurts: the hours lost every week, the process that keeps breaking. The architecture follows from that, never the other way around.
In practice: Mentorway started from coaches juggling Drive, WhatsApp, Calendly and a spreadsheet.
03Booking app for a coworking network
WorkPlus
Book a desk or a room in under a minute.
Visit book.workplus.co.il ↗
01 · Mentorway · How it’s built
One screen for the whole coaching business.
The decision
Tools, not raw access.
The assistant works through 30 defined tools (clients, programs, tasks, calendar, meeting prep) and has no free-form database access. It can do exactly what those tools allow, and nothing else.
- 01An AI agent with 30 tools over clients, programs, tasks and the calendar
- 02Meeting transcripts and AI summaries the moment a call ends
- 03A model layer that switches between AI providers without touching the product
- 04Three languages, including full right-to-left Hebrew
02 · Reelio · How it’s built
An edit suite that runs itself.
A real Reelio output, 14 seconds.
- 1
Photos
The agent drops in the listing photos they already took.
- 2
Vision
A vision model scores every shot and tags the room it shows.
- 3
Planner
A template turns the strongest shots into a timed storyboard.
- 4
Renderer
A deterministic renderer adds the motion, captions, music and voiceover.
The decision
The model plans. The pipeline renders.
The model chooses and orders the shots through tool calls. A deterministic render pipeline turns that plan into video, and every step runs as a background job with retries, so a failed step gets retried instead of shipped.
- 01A render engine: vision scoring, a timeline planner, a video renderer
- 02An AI tool-use loop that directs the whole edit
- 0321 background jobs that queue, retry and notify
- 04A public MCP server, so other AI agents can create listings and generate media
03 · WorkPlus · How it’s built
Booking without calling the front desk.
The decision
Money rules live in the database.
Members can't write their own credit balance, not even with direct access to the API. Credits only move through two database functions that the server alone is allowed to call.
- 01Passwordless sign-in that links each member to their record in the space's management system
- 02Live room availability pulled from that system's API
- 03Credits enforced inside the database, with row-level security and database functions
- 04Hosted checkout with automatic tax invoices, in an installable right-to-left app
About
I’m Alexander, a senior software architect and AI engineer in Israel. I come from years of building systems that have to hold up under load, and I bring that to AI.
When the AI wave hit, I watched businesses buy tools that worked for a month and then cracked, or quietly left their customers’ data open. The problem was never the AI. It was who built it, and how.
So I work like an engineer: understand the business, design the foundation, then let the model do what it’s good at. I learn by building, adopt new tools early, and teach what works to the engineers around me.
AI engineering
- Tool-use agents and orchestration loops
- MCP servers that let other agents use a product
- LLM pipelines across Claude, GPT and Gemini
- Vision and media-generation pipelines
Software architecture
- System and data-model design on Postgres
- Background jobs, queues and retries
- Payments, invoicing and third-party integrations
- TypeScript and Next.js, from database to interface