
Head of Engineering Branding, Wix, @WixEng | Author of Tiebreaker: The Power of Networking, Communities, and Personal Branding (Dec 2021)
Israel
Joined February 2008
- Tweets34.2K
- Following74
- Followers6.5K
- Likes19.6K
Morad Stern ๐ฎ๐ฑ retweeted
For the first time, we're hosting a Claude Community meetup at @Wix Campus, TLV!
Join the meetup on November 2 to examine the decisions behind building with AI: how to define a problem worth solving, judge whether your solution delivers real value, and turn successful experiments into reliable systems. Three talks with practical lessons:
>> Amir Harel (Symphony by Wix) will show how to build the smallest useful version, evaluate it, and turn successful experiments into repeatable results.
>> Guy Aga will walk you through building a real product and business, from market research and product development to launching an online store with Claude Code.
>> Tali Despins will explore how to learn beyond your expertise, work through the discomfort of being a beginner, and combine skills across disciplines.
๐ Wix Campus, Tel Aviv
๐
November 2, 17:00
Limited spots available. Save your spot:
luma.com/claude-9ja6
Morad Stern ๐ฎ๐ฑ retweeted
ืืื ืืืืื ืืืืื ืืืื ืืคื ืืืื ื ืืืืืืชืืช ืฉืืืืขืช ืืืชืื ืงืื ืืืืคืขืื ืชืืื ืืช?
ืืืขื ืืจืงื ืจืืฉ ืืืืืช ืืืืื ืกืืื ืก ื-Wix ืื ืืืฉ ืืช ืืืืืจ ืืืืฉ ืฉืื - From Autocomplete to Intelligence: How Language Became the Operating System of AI - ืืืกืื ืคืชืื ืืขืืจืืช, ืืฉืืื ืฉืืืื ืืืืื ืืืืื.
ืืืืืจ ืฉื ืืืขื ืืกืืืจ ืืื ืืืืื ืฉืคื ืืืืืื, ืืื ืืฉืคื ื ืืฆืืช ืืื ืืืคืืช ื-AI, ืืืืคื ืืืืืืืช ืฉืืื ืขืืืื ืืืืืืืช. ืื ืฆืจืื ืืืข ืืืงืื, ืจืง ืกืงืจื ืืช. ืืืืื ืื ืืงืจืื, ืืฉืืื ืืืืฉืืืจ ืืขืจืืช ืืืกืื.
docs.google.com/document/d/1โฆ
AI can write code, use tools, and power autonomous agents. Yet at its core, it does one surprisingly simple thing: predict ๐ฎ the next word.
So how did autocomplete become intelligence?
In his new article, Gilad Barkan cuts through the weekly noise of new models and frameworks to explain why language became the operating system of AI.
Learn how LLMs actually acquire intelligence, why Transformers and scale changed everything, and where todayโs models still fall short. Along the way, he connects ideas from leading thinkers and AI pioneers, including @ylecun, @AndrewYNg, and @harari_yuval.
No technical background required:
wix.engineering/post/from-auโฆ
Thanks @morad for making good things to happen @ @Wix !
I was happy to share my knowledge and experience from building AI tools (and Octocode), and I feel privileged to be in this position.
For anyone who wants to learn and do the workshop, it's right here:
github.com/bgauryy/octocode-โฆ
Yesterday at @Wix Campus TLV, we got hands-on with Octocode ๐ in a practical workshop led by @guy_bary.
We built an MCP server, created tools AI agents can use, and practiced researching real codebases. We also explored how Octocode connects local code, GitHub, and npm to help agents ground their answers in evidence, alongside best practices for developing reliable tools with MCP and Skills.
This is an important โค๏ธ part of what we do at Wix Engineering: sharing practical knowledge with the developer community and the wider industry. Workshops like this give developers room to experiment, ask questions, and bring what they learn back to their teams - helping them build better and move faster.
Thank you, Guy, and everyone who joined us to learn and build together!
* Slides here >>
github.com/bgauryy/octocode-โฆ
* Octocode here >>
octocode.ai/
Morad Stern ๐ฎ๐ฑ retweeted
ื-AI ืืืืฅ ืืช ืืชืืืช ืืงืื. ืืื ืืฉืืชืื ืื, ืืืืืงืืช ืื-Review ืื ืขืืืืื ืืงืฆื - ืฆืืืืจ ืืืงืืืง ืคืฉืื ืขืืืจ ืืืงืื ืืืจ.
ืืืกืืจืช ืืืืืืคืื ืฉืื ื ืืงืืืืช ืืืคืชืืื, ืฉืืชืคื ื ืืืื ืชืืื ืืช ืืื ืืืช ืงืจืืืืืช ืืืขืืืื ืขื ืืื ื ืืืืืืชืืช. ืขืืฉืื ืจืืืื ื ืืืชื ืืืชืื ืืืืงืืืื - @geektimecoil.
๐งฐ ืืื ืืื ืื ืชืฉืชืืช ืงืื ืืงืกื ืฉืืืืจืช ืกืืื ืื ืืงืื, ืืชืืขืื ืืืืืืข ืืคืจืืืงืฉื, ืืื ืืืื ืืืขืืืก ืืืื ืขื ืืคืจืืืคื ืื ืขื AGENTS .md.
โ๏ธ ืืื ืืขืจืืืื ืกืืื ืฉืื (ืืืื ืืืืื ืืกืืื ื ืืืฉื ื ืฉืื), ืืืืืื ืจืืจืกืืืช ืืฉืื ืืืื ืืขืืืืืช ืืืืื ื ืืจืืฆื.
๐ ๏ธ ืืื ืืืืจืื ืกืืื ืื ืืืืืืช ืืืฉืื ืฉื ื-CI, ืืื ืฉืืืืื ืื ืชื ืืขืืืช, ืืชืงื ืืืืจืืฅ ืืืืงืืช ืืืืฉ.
๐งฐ ืืืื ืืืื ื-xEngineer ืืฉืื ืืืืืืืช ืืืื ื ืขืืืงื ืขื ืืืจืืืช ืืงืฆื ืืงืฆื ืื ืขื ืืจืืืืงืืืจื, ืืื ืขื ืืืืืช ืืชืืฆืื ืืคืจืืืงืฉื.
ืชืฉืขื ืืคืชืืื ืืืคืชืืืช, ืืจืืข ืืืืืืช ืืฉืืืืช ืืฉืืื ืืืช ืงืจืืืืช: ืืื ืืืคืืื ืงืื ืฉื ืืชื ืืืจ ืืืชืจ ืืืืฆืจืื ืฉืืืืขืื ืืคืจืืืงืฉื ืืืจ ืืืชืจ - ืืืื ืืืชืคืฉืจ ืขื ืืืืืช?
ืืืชืื ืืืืื:
geektime.co.il/how-ai-is-chaโฆ
ืขื:
@aviranm, @Asafy3, @guy_bary, @snirshr
Morad Stern ๐ฎ๐ฑ retweeted
From building production AI agents to hiring the engineers who will build what comes next >>> we're proud of @AshkenaziOrtal and Guy Nahum, who are bringing their experience to @Reversim Summit 2026! ๐ ๐
๐ค Ortal will explore what happens when one LLM skill becomes hundreds: how to define skills, catch common failures, and route requests to the right capability in a system handling 300K conversations a month. Her talk builds on research co-authored with Vitalii Kloz and Mykhailo Ulianchenko, published on @arXiv.
๐ค Guy will tackle "Hiring Post-Claude": what makes a great developer today, why junior engineers matter, and how @Wix is rethinking hiring and development programs to cultivate the #xEngineer.
Heading to #Reversim? Don't miss their sessions!
Morad Stern ๐ฎ๐ฑ retweeted
This research by @AshkenaziOrtal, Vitalii Kloz, and Mykhailo Ulianchenko, published on @arxiv - will be featured at the upcoming @reversim 2026 conference! ๐
Don't miss Ortal's session, "From a Single Skill to a Sea of Skills: Architecting LLM Agents in Production".
summit2026.reversim.com/sessโฆ
Really impressive work from @Wix data science team, who published research based on a real challenge they saw and solved in our own customer care agent.
They researched how to help AI agents distinguish between skills that are relevant to a userโs question and those that can actually be executed for that specific user. This helps users get more relevant and useful support from the start.
The team was able to find a more precise approach that reduced the information the model needed to process by 90.5% and saved 228.8 million tokens.
A huge achievement. Congratulations to @AshkenaziOrtal, Vitalii Kloz, Mykhailo Ulianchenko, and the entire team.
Morad Stern ๐ฎ๐ฑ retweeted
1,000+ MySQL servers. 160+ clusters. Multiple regions. Zero downtime.
Join us on October 28 as Shmuel Mekonen and @Danaavital58 break down how they pulled it off - the architecture, automation, tradeoffs, failures, and key decisions:
wix.engineering/migrations-aโฆ
Morad Stern ๐ฎ๐ฑ retweeted
ืืกืืื ื AI ืืืจ ืืฉ ืืืฉื ืืืืืืื, ืืงืื, ืืืกืฃ ืืืืขืจืืืช ืืืืชืืืช. ืืื ืื ืืื ืืื ืืืฆื ืืช ืืืืื, ืืืขื ืื ืืืฆืข ืคืขืืื ืฉืื ืืืงืฉื ื - ืืืฉืืืจ ืืืจืื ื ืืง. ืื ืืืจืื ืขื ืื?
ืืืืืง ืืฉืืื ืืื ื-FTC, ืจืฉืืช ืืกืืจ ืืคืืจืืืช ืฉื ืืจืืดื, ืืชืืืื ืขืืฉืื ืืืชืขืจื. ืืจืฉืืช ืคืชืื ืืืงืืจื ื ืื OpenAI, ืื ืชืจืืคืืง ืืืืจืืช AI ื ืืกืคืืช ืกืืื ืืกืืืื ืื ืฉื ืกืืื ืื ืืืืื ืืืืื.
ืืคื ืจืืืืจืก, ืื ืคืขืืืช ืืืืืคื ืืคืืจืืืช ืืจืืฉืื ื ืืืจืืดื ืฉืืชืืงืืช ืืฉืืจืืช ื-ืดrogue AI agentsืด - ืกืืื ืื ืฉืืืจืืื ืืืืืืืืช ืฉืืืืืจื ืืื.
ืืืฉืืขืืช ืคื ืืื ืืืืื ืืื. ืื ืขื ืขืืฉืื ืืืืืืืช ืฉื ืกืืื ื AI ืืืชื ืืขืืงืจ ืืขืื ืื ืืกืืช. ืขืืฉืื ืืื ืืชืืืื ืืืคืื ืื ืืฉืืื ืฉื ืืืจืืืช ืืฉืคืืืช ืืืืืคื.
ืืื ืฉืื ืื ื ื ืืชื ืื ืืกืืื ื AI ืืืชืจ ืืื ืืคืขืื ืืื, ืืืืืฉ, ืืืจืืฅ ืงืื ืืืืคืขืื ืืขืจืืืช, ืืฉืืื ืืืจ ืื ื ืฉืืจืช ืจืง ืดืืื ืื ืืืืืืด, ืืื ืื ืืืจืื ืืฉืื ืืืขืื.
Morad Stern ๐ฎ๐ฑ retweeted
Bae Code can now work on your own enterprise GitHub repo in the cloud, just like you would work on Base44.
Try it out.
Today weโre introducing Base Code.
Itโs an early preview of how we think software will be built in the future.
Itโs a product that encapsulates everything weโve learnt from:
- How our users are building software.
- How weโre building internally, scaling to hundreds of millions of dollars in revenue while keeping our engineering team very small and focused.
Here are some of the principles behind it, which align with how we at @Base44 think about the future of software engineering:
Cloud:
- Software will move past local desktop environments (where current tools are widely used) and move to the cloud.
- Moving everything to the cloud is not easy. Setting up dev environments with databases, infra components, services, mock data, etc. is easier said than done.
- Base Code first scans your code repo and sets up everything - every infra component (databases, Redis, etc.) - in the cloud. It runs nonstop until your preview environment is ready.
- From an enterprise standpoint, itโs also the logical thing to do: a centralized, governed dev environment instead of handing out keys and secrets to all team members.
- Once Base Code does that, EVERY team member can work in this environment. No more syncing local environments.
Internally, this (cloud) is one of the main things that enabled us to move so fast.
Collaboration:
- Once everything is in the cloud, everybody can write software from anywhere (any browser, your phone, WhatsApp / iMessage coming soon).
- You can easily see what everyone is working on, where theyโre at, and how theyโre prompting - and jump to their environment in one click.
- Weโve built many great collaboration features from the ground up.
Loops, automations, software factories:
- A full, working cloud environment allows for many advanced capabilities we will unveil soon-think agents running in the browser and testing on every device and in any browser.
- It also allows the use of automations to build software factories-e.g., agents reading support tickets, identifying bugs to fix or features to develop, implementing the changes, verifying them in the cloud, and potentially pushing them.
And lastly, thereโs a real advantage in being model agnostic.
--------------
As always, weโre releasing it very early. Itโs far from perfect-but weโre looking for early feedback so we can build it together with our great community and make this vision a reality.
*Weโre giving it away for free for 30 days*.
Give Base Code a try. Connect any GitHub repo, get your live preview, and share it with your team.
Iโd love to hear what works and what doesnโt:
app.base44.com/base-code
Morad Stern ๐ฎ๐ฑ retweeted
ืกืคืืืกืืงืก ๐ ืขืฉืชื ืืืื ืฆืขื ืืฉืื ืงืืืื: ืืฉืืจืจืช ืืืืืื ื Starlink V3 ืืืืืจ ืืืืฉ ืืขืืจืช ืืกืืืจืฉืืค.
ืืื ืื ืืฉืื? ืื ืืคื SpaceX, ืฉืืืืจ ืืื ืฉื ืกืืืจืฉืืค ืืืื ืืงืืช ืืืชื ื-60 ืืืืืื ื V3. ืืืืืจ ืฉืืืืจ ืืื ืฉื Starship ืืืื ืืืืกืืฃ ืืจืฉืช ืงืืืืืช ืฉืฉืืื ืืขืจื ืื-20 ืฉืืืืจื Falcon 9 ืขื ืืืืืื ื V2.
ืฉืืคืืจ ืืืืจ ืืืขืืืืช ืืคืจืืกื ืฉื Starlink.
Morad Stern ๐ฎ๐ฑ retweeted
ืืื ืืืฉืคืช ืืฉืงืคื VR ืืืฉืื ืืืฉืงื 100 ืืจื ืืืื - ืืืฉ ืขืื (ืคืืืก ืขืืืื ืืฉืื ๐ฎ๐ฑ ืขื ืืฉืจืื!)
ืืื ืก Meta Connect 2026 ืืื ืืฆืืื ืืช Meta VR Glasses, ืืฉืงืคื VR ืงืืื ืืื, ืืฉืืืขืื ืืืกืืืื ื ืืฆืืื ืืืืืืช ืืืฉืื ื ืคืจืืช ืืืืก.ย
ืื ืืืืืื ืืกืื micro-OLED ืืจืืืืืฆืืืช 5K, ืฉืืืื ืืืืฆืขืืช ืืขืื ืืื ืืืืืืื. ืืืืืจ: 1,299 ืืืืจ.
* ืืขืจื: ืื ืืืืข ืื ืื ืชืืื ืืฆืืื, ืืื ืื ืืืืืง ืื ืฉืืคื ืืื ืฆืจืืืื ืืืฆืื. ืคืกืคืืก ืฉื ืืคื.
ืืืงืืื ืื ืืืจืืื ืขืย Ray-Ban Meta Audioย ืืฉืงืคืืื ืืื ืืฆืืื (*ืฉืื ืืขืจื: ืืืื ืืืืฉื ืืฉืืจ ืืืืงืืจืช ืขื ืคืืืขื ืงืฉื ืืคืจืืืืช), ืืืืจ ืืืฉ ืฉื Ray-Ban Meta ืขื ืฆืืืื 3K ืืกืืืื ืืฉืืคืจืช.
ืกืืื ื-AI ืฉื ืืื Muse ืืงืื ืฉืืืืช ืงืืืืืช, ืืืฆืืข ืืฉืืืืช ืืจืงืข, ืืื ืืืจืฆืืืช ืขื ืฉืืจืืชื ืฆื ืืณ - ืืื ืืืืืช ืืขืืื ืืจื ืืืฉืงืคืืื.
ืืื ืื ืืฆืืื ืืช Muse Charmย - ืืืฉืืจ AI ืงืื ืืืืก ืฉืืืคืฉืจ ืืืืจ ืืฉืืจืืช ืขื ืืกืืื ืืื ืืืืฆืื ืืช ืืืืคืื. (* ืฉืื ืืขืจื: ืื ืื ืฉ OpenAI ืืฆืืื ืืงืจืื. ืืื ืืืืจืื ืืขืงืคื ืืืชื)
ืืืฉ ืื ืืงืฉืจ ืืฉืื ๐ฎ๐ฑ ืืืฉืจืื: ืืฉืงืคื Ray-Ban Meta ื-Oakley Meta ืืืืขื ืจืฉืืืช ืืืฉืจืื - ืขืื ื 2026. ืื ืขืื ืจืืข. ืืืืจ ืขื ืืื ืืคืจืืืืช - ืืจืฉืืืืช ืืืืืืช ืืืืืจ ืืช ืื.
Morad Stern ๐ฎ๐ฑ retweeted
ืดืืฉ ืื ืขืื ืืืจ ืืื ืืืฆืื ืืืืด (ืขื ืืฉืงื Oh, one more thing ืฉื ืืณืืืก) ืฆืืงืจืืจื ืื ืืฆืื ืืช Muse Charm: ืืืฉืืจ AI ืืขืืจ, ืืขืื ืืืืืืฆื ืฉืืืืจ ืืชืช ืืืฉื ืืืืจื ืืกืืื ืืืืฉื Muse ืืืื ืืืฉืช ืืืืคืื.
ืืืืฉืืจ ืืืื ืืกื ืืืข ืงืื, ืืืงืจืืคืื ืื, ืืืืืจ 5G ืืืืืฉื ืืืืขืช ืืฆืืข. ืืืืฆืื, ืืืืจืื ืขื ืืกืืื, ืืขื ืืืกื ืืืคืืขื ืืืืช ืื ืืืืืืืช ืฉืืืืฆืืช ืืืชื. ืืขืืจ ืืฉืืืืช ืืชืฉืืืืช, ืืจืขืืื ืืื ืฉ-Muse ืืืื ืื ืืืฆืข ืคืขืืืืช ืืื ืื ืื ืืฉืืืืช, ืืชืื ื ืืืจืื ืืืขืืืจ ืืงื ืืืช ืืืืืื ืืช. ืืืจืง ืืกืืืจ ืฉืื ืื ืืืืจื ืืืื ืื, ืืืย ืืชืื ื ืื ืืืชืืื ืืฉืืืง ืืืชื ืืงืจืืช ืืฆืืืจ.
ืื ืืื ืฆืจืื ืืืืืช ืืจืืข ืฉื OpenAI ืื ืืื.
ืืื ืืืฉืคืช ืืฉืงืคื VR ืืืฉืื ืืืฉืงื 100 ืืจื ืืืื - ืืืฉ ืขืื (ืคืืืก ืขืืืื ืืฉืื ๐ฎ๐ฑ ืขื ืืฉืจืื!)
ืืื ืก Meta Connect 2026 ืืื ืืฆืืื ืืช Meta VR Glasses, ืืฉืงืคื VR ืงืืื ืืื, ืืฉืืืขืื ืืืกืืืื ื ืืฆืืื ืืืืืืช ืืืฉืื ื ืคืจืืช ืืืืก.ย
ืื ืืืืืื ืืกืื micro-OLED ืืจืืืืืฆืืืช 5K, ืฉืืืื ืืืืฆืขืืช ืืขืื ืืื ืืืืืืื. ืืืืืจ: 1,299 ืืืืจ.
* ืืขืจื: ืื ืืืืข ืื ืื ืชืืื ืืฆืืื, ืืื ืื ืืืืืง ืื ืฉืืคื ืืื ืฆืจืืืื ืืืฆืื. ืคืกืคืืก ืฉื ืืคื.
ืืืงืืื ืื ืืืจืืื ืขืย Ray-Ban Meta Audioย ืืฉืงืคืืื ืืื ืืฆืืื (*ืฉืื ืืขืจื: ืืืื ืืืืฉื ืืฉืืจ ืืืืงืืจืช ืขื ืคืืืขื ืงืฉื ืืคืจืืืืช), ืืืืจ ืืืฉ ืฉื Ray-Ban Meta ืขื ืฆืืืื 3K ืืกืืืื ืืฉืืคืจืช.
ืกืืื ื-AI ืฉื ืืื Muse ืืงืื ืฉืืืืช ืงืืืืืช, ืืืฆืืข ืืฉืืืืช ืืจืงืข, ืืื ืืืจืฆืืืช ืขื ืฉืืจืืชื ืฆื ืืณ - ืืื ืืืืืช ืืขืืื ืืจื ืืืฉืงืคืืื.
ืืื ืื ืืฆืืื ืืช Muse Charmย - ืืืฉืืจ AI ืงืื ืืืืก ืฉืืืคืฉืจ ืืืืจ ืืฉืืจืืช ืขื ืืกืืื ืืื ืืืืฆืื ืืช ืืืืคืื. (* ืฉืื ืืขืจื: ืื ืื ืฉ OpenAI ืืฆืืื ืืงืจืื. ืืื ืืืืจืื ืืขืงืคื ืืืชื)
ืืืฉ ืื ืืงืฉืจ ืืฉืื ๐ฎ๐ฑ ืืืฉืจืื: ืืฉืงืคื Ray-Ban Meta ื-Oakley Meta ืืืืขื ืจืฉืืืช ืืืฉืจืื - ืขืื ื 2026. ืื ืขืื ืจืืข. ืืืืจ ืขื ืืื ืืคืจืืืืช - ืืจืฉืืืืช ืืืืืืช ืืืืืจ ืืช ืื.
Morad Stern ๐ฎ๐ฑ retweeted
You know Wix Engineering on LinkedIn. Now come meet us on Instagram ๐
Follow Wix Engineering for the code, the people, the lessons learned, and everything engineering:
instagram.com/wix_eng/
Morad Stern ๐ฎ๐ฑ retweeted
AI context engineering, 756K+ production messages, 6B+ daily media requests, zero-downtime database migrations, infrastructure migrations across 1,000+ MySQL servers, our upcoming meetup and more!
Hereโs whatโs inside our latest @Wix Engineering newsletter -
๐ง From Autocomplete to Intelligence: How Language Became the Operating System of AI / @GiladBarkan
AI can write code, use tools, and power autonomous agents - but underneath it all, it still predicts the next word. Explore how language became the operating system of AI, why Transformers and scale changed everything, and where todayโs models still fall short.
๐ฆ Donโt Offer What Canโt Be Done: Deterministic Executability Gating for LLM Skill Selection at Scale / @AshkenaziOrtal, Vitalii Kloz & Mykhailo Ulianchenko
Semantic matching alone isnโt enough in production. Across 756K+ messages, Wixโs executability gate filtered out 59.4% of matched skills, reduced skill context size by 90.5%, and saved 228.8M tokens by checking what users can actually do before skills reach the LLM. Now as a new paper on @arXiv.
๐ ๏ธ [Upcoming Meetup] Migrations at Scale: Code, Infrastructure & Automation / @Danaavital58 & Shmuel Mekonen
Two migrations, two very different challenges: automating code changes across hundreds of codebases, and migrating 1,000+ MySQL servers across 160+ clusters in 30 days - with zero downtime. [Oct 28, at Wix Campus TLV]
๐ผ๏ธ Architecting for 6 Billion Daily Requests: Inside Wixโs Media Platform / Tomer Salton
What happens when image resizing needs to work across 300M+ websites and more than 6 billion media requests every day? Go behind the architecture powering real-time AVIF encoding, smarter caching, lower resize costs, and secure asset delivery at massive scale.
๐ฏ Attention is Expensive: Optimizing Context Windows for Precision and Cost / Maya Halevy
More context doesnโt always mean better AI. Learn how attention limits, token waste, and the โLost in the Middleโ problem impact LLM accuracy - and how better context design can improve reliability while reducing cost.
๐๏ธ How We Built a Zero-Downtime Database Migration Service at Wix / Bar Shauli
How do you move critical production databases between clusters without users noticing? Explore the Kafka, ProxySQL, and MySQL architecture behind Wixโs internal migration service - from streaming data to cutting over live traffic with zero downtime.
Read it here ๐
wix.engineering/so/2dQ3DoCRl
Really impressive work from @Wix data science team, who published research based on a real challenge they saw and solved in our own customer care agent.
They researched how to help AI agents distinguish between skills that are relevant to a userโs question and those that can actually be executed for that specific user. This helps users get more relevant and useful support from the start.
The team was able to find a more precise approach that reduced the information the model needed to process by 90.5% and saved 228.8 million tokens.
A huge achievement. Congratulations to @AshkenaziOrtal, Vitalii Kloz, Mykhailo Ulianchenko, and the entire team.
[new @arxiv paper] Thereโs a huge gap between playing with AI skills in a demo and โ๏ธ engineering them for production.
In a simple setup, semantic routing feels like enough: the user asks something, the LLM matches the text to a skill description, and executes it. But in real-world production at scale (like @Wix Customer Care), semantic matching alone isn't enough. ๐ฉ
Why? Because semantic matching only looks at what the user says, completely ignoring what the user is actually allowed or able to do.
A user asking to "connect a domain" is a 100% semantic match for the Domain Skill. But if they don't have an active premium plan, executing that skill is a dead end.
In a new arXiv paper by @AshkenaziOrtal, Vitalii Kloz, and Mykhailo Ulianchenko, we present Helpmateโs (Wixโs Customer Care AI agent) architecture: introducing a deterministic executability gate that evaluates account state before candidate skills reach the LLM.
Tested across 756K+ user messages in production:
๐ฏ 59.4% of semantically matched candidates were filtered out because the account state made them non-executable
๐ 90.5% reduction in skill context size 228.8M tokens saved
๐ก๏ธ In a targeted replay of 1,000 risky conversations, removing the gate caused the model to select a non-executable skill in 7.8% of cases.
Read the full paper: arxiv.org/html/2608.01050v1
Morad Stern ๐ฎ๐ฑ retweeted
[new @arxiv paper] Thereโs a huge gap between playing with AI skills in a demo and โ๏ธ engineering them for production.
In a simple setup, semantic routing feels like enough: the user asks something, the LLM matches the text to a skill description, and executes it. But in real-world production at scale (like @Wix Customer Care), semantic matching alone isn't enough. ๐ฉ
Why? Because semantic matching only looks at what the user says, completely ignoring what the user is actually allowed or able to do.
A user asking to "connect a domain" is a 100% semantic match for the Domain Skill. But if they don't have an active premium plan, executing that skill is a dead end.
In a new arXiv paper by @AshkenaziOrtal, Vitalii Kloz, and Mykhailo Ulianchenko, we present Helpmateโs (Wixโs Customer Care AI agent) architecture: introducing a deterministic executability gate that evaluates account state before candidate skills reach the LLM.
Tested across 756K+ user messages in production:
๐ฏ 59.4% of semantically matched candidates were filtered out because the account state made them non-executable
๐ 90.5% reduction in skill context size 228.8M tokens saved
๐ก๏ธ In a targeted replay of 1,000 risky conversations, removing the gate caused the model to select a non-executable skill in 7.8% of cases.
Read the full paper: arxiv.org/html/2608.01050v1
Morad Stern ๐ฎ๐ฑ retweeted
Thank you @HebrewU_heb for highlighting Shmuel Mekonen's work, article and upcoming meetup:
1 - @HebrewU_heb - alumni.huji.ac.il/event/wix_โฆ
2 - Article: 1,000 Servers. 160 Clusters. 30 Days. Zero Downtime - Migrating @Wix's MySQL Fleet to Graviton
wix.engineering/post/1-000-sโฆ
3 - Meetup: Migrations at Scale: Code, Infrastructure & Automation
wix.engineering/migrations-aโฆ
[Meetup] Large-scale migrations are never just about moving code or infrastructure. They're also about doing it without ๐ฉ slowing everyone else down.
At @Wix, two very different migration challenges required two very different approaches:
1 - Rolling out automated code changes across hundreds of codebases, while keeping developers focused on day-to-day work.
2 - Moving 1,000+ MySQL servers across 160+ clusters, multiple regions, and two major platform transitions - all within 30 days and with zero downtime.
Dana Avital and Shmuel Mekonen will share how these migrations were designed and executed: the architecture behind them, the automation that made them possible, the tradeoffs, the failures along the way, and the key decisions that helped them succeed.
Join us on October 28 at 17:30 at Wix Campus TLV:
wix.engineering/migrations-aโฆ
Morad Stern ๐ฎ๐ฑ retweeted
ืื ื ืืืื ืืขืืื ืขื ืืืง ืืืฉืื ๐ค ืืืืืงื, ืืื ืชืืื ืืคืจืืข ืื ืื ืชืง ืืื: ืืืืืืงื ืืืืื ืืืฉืชืืื, ืืื ืืืกื ืืืฉืืจ ืืืืฉ ืืืืจื.
ืื ืื ืืชื ืืคืืืงืฆืืืช macOS ืืืฉื ืฉืืืืื ื ืืกืืื ื ืฉืืชื ืื ืืืืฉื, ืืืคื ืืืืืืง ืืกืคืืืืคืื ืืขื ืืืืืื ืืืคืืคื - ืืืืคืืช ืืืชื ืืืคืงืืื ืืืื โจ ืฉืืืงืขืื ืืงืฆืืืช ืืืกื.
ืฆืืขืื, ืืืงืืงืื ืืชื ืืขื ืฉืืืืืื ืืืื ืืืช ืืืืืืงื ืืืงืฆื, ืืืื ืืืกืชืืจ ืืช ืืชืืื ืฉืื ื ืขืืื ืขืืื. ืคืชืืื ืืืืฉื ืฉืื ืื ืจืง ืืฉืืืข ืืช ืืืืืืงื, ืืื ืดืืจืืืฉืด ืืืชื ืืืชื!
ืืืคืืืงืฆืื ืชืขืื ืืงืจืื ื-App Store. ืื ืชืจืฆื ืฉืืขืืื ืืชืื ืืฉืืื ืืืืืืจ ืฉืืื ืื ืืืืขื.
Morad Stern ๐ฎ๐ฑ retweeted
ืืคืื ื ืืืืช ืกืคืจ, ืืฉืจืื ืื ืฉืฉื ืงืืื ืื, ืืฉืจ ืฉืืืจืื ืขื ืืขืจืืช ืืืืคืช ืืืืง ืืืืง ืืืืืง ืืคืชืืืช ืฉืืื ืคืืชืื ืืืฉืงืืืช ืขื ืืืืืืื ืืืฆืืจืช ืืงืื. ืืื ืืื ืขืืจืืื ื - ืงืืืื ืืืืืชื.
ืืื ืืืฉืื? ืจืง ืขื ืื ืฉืืื ื ืืช ืืืื ืืืกืืื ืืื.





















