Backend & Platform · Distributed Systems · EU Remote
10+ years on systems that must work in production — Shopify logistics (Staff), Novu (employed), Sedric platform/DX. Founder of ZaZet Solutions (SMB automation, HR / EN / HE). Open to part-time Staff/Senior backend & platform roles. Available for consulting on AI-assisted dev workflows and organizational DX.
SMB automation and custom operational systems (booking, payroll, internal tools). Same platform discipline as employed roles — teams of 5–50. Open to part-time Staff/Senior backend roles in parallel.
Unified developer data platform (React, TypeScript) consolidating AWS + MongoDB Atlas for CSM and engineering. Org-wide AI-assisted delivery with Claude Code — CI overhauls, tooling, workshops.
Full-time on OSS notification platform (35k+ repo stars). Cut build time 78%; redesigned Novu V2 API; OpenAPI in CI/CD; channel CLI generators.
Commerce logistics post-acquisition. Cut operational delays 70%; shipment events + audit log → 97% on-time shipping; cross-team integration on Java/TypeScript microservices.
B2B product 0→1 before Shopify acquisition. OpenAPI adoption across 8 teams / 14 services — API overhead ↓90%. Warehouse back-office chatbot tooling.
Vaadin back-office adopted day-one by CSM. Zero-downtime Elasticsearch migration — 46% cost ↓. Release turnaround ↓50% via E2E practice.
Enterprise PLM — Java backend, manufacturing lifecycle tools, modern JS in legacy systems.
First engineering role — production web apps across Germany and Austria.
Verified outcomes — see employment schedule for context.
| Value | Metric | Source |
|---|---|---|
| 78% | Build time reduction | Novu |
| 97% | On-time shipping | Shopify |
| 90% | API overhead reduction | Deliverr |
| 70% | Operational delays reduction | Shopify |
| 46% | Infrastructure cost reduction | Wochit |
| 10+ | Years backend/platform | — |
Apr 2024 – Jun 2025 on the notification infrastructure platform (35k+ repo stars). V2 API architecture, OpenAPI-driven CI/CD, channel CLI generators, build-system optimizations.
Notifications died, calls kept ringing. Two weeks through every iOS setting led to my own self-hosted WhatsApp bridge, quietly announcing me online 24/7 - traced into a one-line assumption bug in Baileys itself: a bare presence node sent on every incoming message, even after the documented fix. Forked it, added a regression test, filed it upstream. The identical fix was proposed once before and closed by a stale bot.
The manual for the fork. The licence question answered first - public fork under AGPL and you are fine, private modifications and you are the case the Enterprise licence exists for. Then every feature on the branch (RTL, form mode with repeating groups, recipient upload slots packaged into the sealed PDF), both API surfaces with the full v1 route list, the Playwright HTML-to-PDF path, and the two things you must change in the stock docker-compose.
DocuSign costs per envelope, DocuSeal puts the API behind the paid tier, and Documenso sealed my Hebrew as empty squares. So I forked it: real RTL, a Hebrew font inside the PDF sealing engine, and a form-style signing mode where the signer fills a labeled Hebrew form with the document updating live beside it. Proven on Tofes 101 - all 173 fields, none of them placed by hand: code reads the PDF text layer and measures, the agent looks at the render and corrects. 0.14% off. Live demo mints you your own private copy.
My agent got a real Chrome in the cloud - and a way to hand it back. When it hits a password, a 2FA code or a Pay button, it sends its browser live to my phone, waits (a verified 925 seconds), and carries on with the login intact. It never sees a credential. Also: the free-tier quota that impersonated a keep-alive bug, the panel that covered every login form, and the Done button nobody needed.
Every meeting-bot tool wants to join your call as a participant and bill you monthly. Mine records whatever is making sound on the machine - WhatsApp, Meet, Zoom, a phone on speaker - with BlackHole and one ffmpeg command, transcribes locally with whisper.cpp, and drops a plain .txt transcript in ~/Calls you can pipe into Claude Code, an LLM, or grep. Now with a Windows build that needs no audio driver at all, and hasn't been tested yet.
Four household finance audits done by hand, turned into a tool that runs on your own machine - built on Eshed Shaham's israeli-bank-scrapers and Brafdlog's Caspion, because Israel sits outside PSD2 and no aggregator API exists. Why the regulation chose the architecture instead of killing the idea, the four real bugs behind it (including a 448KB Chrome that exits 0), and the checklist that is the actual product: six ways to be confidently wrong, four of which pass an arithmetic review. Beta testers wanted.
A credit card revolving at 17.69% since 2006. If nobody ever added a shekel, ILS 232 back then is ILS 6,028 today - doubling every 4.3 years, with no schedule and no date it ends. Five structural reasons nobody ever sees it, plus what else a year of real bank data held: idle cash a bank away from its own debt, PayPal billing agreements that renew with no email, and why cancelling a "duplicate" סיעודי policy loses you money.
I sat down to fix his Start button and found four antivirus products fighting each other, five weekly app subscriptions (the same wallet app twice, at two different prices), and a third of his phone bill that turned out to be kitchen appliances at three times retail. The mailbox sweep, the PDF parser, and the refund argument that got the first one approved two days after the window closed.
Read → · field note
Every client opens me an internal email so the work stays in their scope - GitHub, Supabase, Base44. The browser was the last hole: a task space is born on whichever profile is active, and the option meant to control it was accepted and silently dropped. The fix is an email as the handle, a profile map that builds itself, and a hook that refuses any run that does not name its scope.
Five subagents dug through 8 months of my own logs: May cost me 9M tokens per session, August 1.7M - same workload, cache ratio unchanged. The habits that made the difference, and a skill (claude-setup-audit) that grades your setup against them from your local logs. I got a C.
The comments were better than the post. One question sent me back to 58,617 turns of raw logs, where I found the thing I had missed: a token added early in a long session is re-read once per remaining turn, 992 times on average. One session I forgot to close ran 24 days and was 17% of my month. The script that grades you is in the post.
Read → · The original · repo ↗
The comments on the ego-lite post turned into a shopping list, so I sent five research agents in parallel with one fixed six-question sheet. One tool dies in six days, one was steered into a 1Password vault, and only three even play the same game.
Twenty-six agent panes across seven workspaces, and macOS still just says "Claude is waiting for your input". cmux is a free Ghostty-based terminal that knows an agent is running inside it. Here is my full cmux.json - including the discovery that three quarters of my "tuned" config was the defaults, and why 26 agents only fit in 24 GB because I have not opened PyCharm since June.
People burn through a Max plan in days; I put 13.9 billion tokens through mine in two months and still cannot hit the limit. I opened my own logs to find out why. 96.3% of my tokens are cache reads, cached tokens draw your limit at the cached rate, and the turn that actually burns your plan looks like a one-line question.
For weeks I typed /remote-control at the start of every session so I could approve prompts from my phone - until I forgot once and left the house. One line in settings.json makes every session boot already connected, and it quietly changed when I start work.
A reader asked the one question I could not answer, so I measured it: four of five impossible browser actions came back as success. The exact boundary between an error and a silent no-op, the phantom edit that is worse than both, and the one line that catches all of it.
Browser automation makes you choose between an extension that steals your Chrome and a headless browser logged into nothing. The third option: an isolated task space that inherits your logins - and a control model where the user always wins.
I exposed the HDR glow trick last week. Then someone said that brightness physically hurts their eyes. So I went the other way — into macOS's private CoreDisplay APIs — and found why killing the burn locks your brightness slider, and how to get both back. Shipped as GlareBar, a ~150-line menu-bar app that kills HDR glare and dims below the hardware floor.
Browser automation for Gmail is flaky; there's an API. 260 lines of Python gives an agent search, read, attachments, threaded replies and labels, then the browser handles only what email can't. One real flight check-in, three catches a human would have missed.
That glowing Wiz logo on LinkedIn wasn't a filter — it was an HDR colour profile (Rec.2020 + PQ) making its white pixels decode at ~4,000 nits. I reverse-engineered it, rebuilt it as a one-command tool + Claude skill, and wrote up the half of the story nobody tells: on non-HDR screens the same file looks broken.
caffeinate keeps the screen on but ignores a closed lid. So I had Claude Code build a menu-bar toggle with three honest states — an in-process IOPMAssertion, a sudoers rule scoped to two commands, ~90 lines of Swift, no dependencies.
Four agents on a EUR 6 VPS, reachable by Telegram and email. What NanoClaw is, the three integration bugs that fail silently, why the personas are the product, and the trifecta I deliberately did not close.
Why terminals render Hebrew backwards (the Unicode BiDi algorithm), how to fix it on WezTerm, and rebuilding a cmux-style multi-agent Claude Code workflow on top — installed by the agent itself.
I help engineering leaders move from “everyone experiments with AI in isolation” to a repeatable delivery system — the same work I drove at Sedric: platform tooling, CI modernization, hands-on enablement, and measurable velocity gains without sacrificing quality.
Reference implementation: Sedric.ai — org-wide AI-assisted delivery, unified dev data platform (AWS + MongoDB Atlas), CI overhauls, workshop-led rollout. Engagements are remote-friendly, typically part-time or fixed-scope.
gal@tidhar.org.il — subject: Consulting — AI workflows & DX
Production systems for SMB operators.
4-language production system. Replaced phone/spreadsheet booking.
30-person company — payroll days → hours.
One custom ops system replacing department spreadsheets.
Platform thinking from Sedric and Shopify, applied to teams of 5–50 — booking, payroll, ops without generic SaaS bloat.
zazet-solutions.hr — book intro call
Pick your path — recruiting, consulting, engineering, or business.