Keeps a twelve-site cinema chain's IT and security running, through every major switchover. Founder and product manager of PERFIX, a home-services marketplace with 3,402 customers across 45 towns.
What the estate couldn't do and what it does now
- The sales platform ran on local servers and end-of-life operating systems, at every site.
- It moved to a cloud platform. About 400 terminals were re-imaged and redeployed across 12 sites, and the cinemas kept selling throughout.
- Security regulation said every endpoint had to leave its end-of-life system at once, and an upgrade can lose data, programs and the special programs a cinema runs on.
- I built an upgrade automation that runs on each machine: it fetches the right Microsoft image for that specific endpoint, upgrades silently without losing anything, then strips the apps Microsoft pushes in, applies the group policies and locks the machine down. Two rule sets: an employee PC and a point of sale are not the same machine. The upgrade is the easy part; keeping every program and every file is the job.
- A full room of physical servers, one box per job, with constant issues across tens of machines.
- The company servers went virtual, on our own hosts. One place to maintain instead of tens of boxes, with backups that go back weeks and a disaster-recovery copy.
- HQ was one flat network: printers, cameras, endpoints, everything together.
- It was split into VLANs with explicit rules for each, where it may go and where it may not, and every computer is registered by its hardware address. Plug an unknown machine into a switch and it gets no address until it is onboarded.
- The old endpoint protection let things slip through.
- The estate moved to CrowdStrike, and what used to slip through gets caught straight away. A strong product, but only if someone manages it properly. That's me.
- Every audit and penetration test comes back with a list.
- MFA everywhere, domain rules and group policy, the virtualization hosts upgraded, the right backup retention, 2007-era servers replaced with ones that aren't end of life. Regulation sets the list; the job is to get through it in the fewest steps, on many machines at once, without losing anything. Generic tools rarely fit a cinema chain, so I build the company-specific one.
What the first version taught and what the rebuild does about it
- The deal left the app after the first chat.
- PERFIX became the seller: one price to the customer, and PERFIX pays the provider. A seller can guarantee the job; a middleman can only apologise.
- Customers wrote "it doesn't work", and nobody can price that.
- A catalog of standard jobs: the provider names the work, PERFIX sets one price before anyone arrives, and an unclear job starts with a paid diagnosis visit.
- I was the matching system, by hand.
- Matching runs on a reliability score built from each provider's recent jobs. Star ratings feed it, but nobody is ranked on an average of stars.
- More deals died on reaching each other than on price.
- In-app chat and masked calls during the job: they can talk, but neither ever sees the other's number.
- "Everything included" starts arguments.
- A written scope of what is in and out, and an open dispute pauses the provider's payout until it is settled.
- People paid cash, one deal at a time.
- Card payments and installments. The goal isn't cheaper, it's payable.
- The busiest category was the worst served.
- The plan's first technician is an appliance specialist on the payroll: quick fixes by PERFIX's own people, licensed work and projects by vetted partners.
Earlier how it started
Projection & sound · Hot Cinema, 2021–2022. Dante-certified audio over IP. Ran the projection and calibrated the chain's auditoriums (5.1 and 7.1), set up the sound systems for new auditoriums, then the technical team, which at a cinema chain is also the IT team. Christie projectors, Dolby cinema servers and sound processors, about 400 digital-signage screens and the video walls.
Sound engineer · Porti Records, Tel Aviv, 2021–2022. Managed and operated recording sessions, in parallel with the cinema.
The cast tools and skills
- Identity & Microsoft 365
- Microsoft 365 · Exchange Online · Entra ID · Active Directory · Group Policy · Conditional Access · MFA (FortiToken)
- Security
- FortiGate (VDOM, VLANs, policies) · CrowdStrike EDR · Action1 · Cloudflare Zero Trust · phishing simulations · audit remediation
- Network
- VLANs · site-to-site and remote VPN · DNS · DHCP · TCP/IP · Wi-Fi with Ekahau site surveys
- Infrastructure
- Windows Server · VMware ESXi · Hyper-V · Veeam · Acronis · QNAP · backup and disaster recovery
- Cinema & retail tech
- ~400 POS terminals · Christie · Dolby · Screenwriter TMS · Samsung MagicINFO · video walls · event streaming
- Product
- PRDs and specs · user flows · Figma and FigJam · wireframes · funnel instrumentation · unit economics · GA4 · Meta Ads · Jira · Notion
- Cloud & build
- Cloudflare (Pages, Workers, Tunnels, Access, WAF) · AWS · Firebase / Firestore · MongoDB Atlas · Tranzila payments · InforU SMS · Codemagic · TestFlight
- Automation & scripting
- PowerShell · Bash · Python · SQL · n8n (~63 workflows) · Git · Postman · REST
- AI systems
- Retrieval over private documents · multi-agent pipelines · local open-weight models · prompt and cost governance · Claude and Claude Code
Behind the scenes Command Desk, the control room I built
Command Desk is an operations control room, web and mobile, shipped to TestFlight on every push. It reads my notes and the live design files, drives the testing, holds the tasks, and routes each piece of work to the AI session that already knows the context. Around it runs one loop that keeps the designer, the developers and me on one truth. I built it, nights and weekends, while the product was being built.
- 1 · Wireframe
- Every screen of the customer app, the provider app and the admin console is drawn before the designer starts, and handed over as a brief with every frame's ID. Every wireframe and every designer frame is extracted into a searchable catalog, refreshed daily, so a question about the design is answered in seconds.
- 2 · Design check
- When the designer delivers, a script parses both Figma files into screen inventories, matches every screen to its wireframe and compares every visible text, control and link (headings, actions, table values, money amounts). The designer gets a ranked fix list (high, medium, low) and the fix notes are placed on the frames themselves. Last round: 84 findings on 46 screens, down from 162 on 38 the round before.
- 3 · Handoff
- The developers get one page: a clickable prototype of the admin console (98 screens, 4,097 live links, nothing dangling, 42 operator-task tours), a 431-page developer guide with one section per screen (purpose, data, formulas, controls, rules, states), a formulas sheet, a navigation map of every click on every screen (2,263 rows), and 63 Jira tickets generated from it, each linked to its screen and its guide section.
- 4 · Build test
- Each release goes into two fresh emulators. A driver steps through both apps and screenshots every step, with read-only peeks into the database and the admin console behind each one. Every designed screen (134 customer, 127 provider, 91 admin) is marked built as designed, built differently, dummy or missing, with its screenshot. The register and a slide deck built from it run the meeting with the dev team. A separate QA agent can be pointed at any app, web or Android, and reports what is broken and what is missing, ranked, with evidence. "Inconclusive" is never a pass.
- 5 · Fix
- The findings go back where each person works: Jira tickets for the developers, fix notes in Figma for the designer, and the catalog itself goes live from a spreadsheet (12 trades, 124 subcategories, 188 priced units of work), checked and applied to the live database by a script, each icon approved on a Figma board before it enters the app.
- The one law
- The machine recommends; I confirm. Every automated step ends at a gate where I say yes. Behind the desk: a private search engine over my notes, specs and books on my own hardware, a multi-agent pipeline with each agent limited to the tools it needs, and open-weight models on my own GPU. Built with AI-assisted development: I design the system, set the rules and run it.