Who
AI product builder. Former Israeli Air Force air control officer, Major (res.), then project manager and strategic advisor at the National Drone Initiative (2024–2026). Lives in Tel Aviv.
Ori Ringel · AI Innovation Builder · Tel Aviv
I used to direct aircraft.
Now I direct AI agents.
For eight years in the Israeli Air Force I directed aircraft from an underground control unit. I never saw them; I worked from radar, radio and read-backs. AI agents are the same: you can't watch them think, so you build the instruments and check the evidence. That's how I take ideas from “what if” to shipped products.
- 2.2M YouTube views
- 17 days scaffold → App Store
- 2,777 automated tests
- 10 personalised books live
SECTOR ORI · 4 TRACKS · 1 INBOUND
Each blip is a project. Select one to open it.
00 · In 30 seconds
The 30-second version.
Proof
- MikoMikoAI kids' music videos + a personalised book store · 2.2M views
- NutsHebrew/English learning app, live on iOS and Android
- WandyAgents that build and verify automations · pilot
- 3D Model StudioFrom a reference image to a printable, tested 3D file
Foundation
15 months of software-development training at Sela College, the Deep Learning Specialization (Andrew Ng) and Mathematics for Machine Learning: Linear Algebra (Imperial College London). I read and write code, and I learn by shipping.
Why eTeacher
- Real learners50,000+ students a year (per your job post), so what I build gets used and measured.
- Learning productsWhat I most enjoy building: Nuts teaches why a decision is right, not just whether it was.
- HebrewMy native language, and I've already worked through hard Hebrew-AI problems. You run a Hebrew school.
01 · Built for eTeacher
I didn't just write a cover letter.
I built you something.
In August, eTeacher's CEO wrote that AI that performs for people is not the same as AI that develops them ↗, and that when the tool is removed, the gains often vanish. When I trained new air controllers, the test was the same: could a trainee hold the picture once the instructor stepped back? So I built a small, working version of that idea for the kind of learner your Rosen School of Hebrew teaches.
Prototype · built for this application
The Practice Room: a hint ladder and a tool-off test.
A between-classes practice set that asks before it answers, then switches the hints off to measure what the learner can do alone. Play it: the English is under every line, so no Hebrew is needed.
A prototype, not connected to eTeacher systems. I wrote the content and it would need a teacher's review. The hints are hand-written in this v1; a real version would draw them from approved course material. Audio uses your device's Hebrew voice if it has one.
02 · Projects
Four systems, built end to end.
Each one started as a question I couldn't stop thinking about. I own the idea, the product decisions, the architecture and the review. AI writes much of the code and media under my direction, and I'm explicit below about which is which.
- MIKO01 MikoMikoAI kids' music studio + personalised books STATUSLIVE METRIC2.2M views
- NUTS02 NutsBilingual poker-learning app STATUSLIVE METRICiOS + Android
- WNDY03 WandyAgents that build and verify automations STATUSPILOT METRIC380 tests
- 3DMS04 3D Model StudioReference image → printable, measured file STATUSWORKING METRIC6 tested models
MIKO01LiveYouTube channel + store
MikoMiko
An AI production studio for Hebrew kids' music videos, connected to a store that makes each child the hero of their own illustrated book.
- The question
- Can one person directing specialist AI agents produce Hebrew children's content that kids actually watch, and turn that attention into a personal product?
- What I built
- Two connected systems. A production line where agents take a brief to a published music video. And a personalised-book pipeline: each story's art is generated once with Higgsfield MCP, the child's photo, uploaded by the parent, is checked by Google's Gemini API, and every page is redrawn with that child as the hero. Watch it step by step →
- Result
- 2.2M+ views by Oct 2026, from a channel started in Oct 2025. The 29 Jun 2026 analytics report: 1,833,018 views, 99,645 watch hours, 23 videos. 10 personalised book titles live at mikobooks.store ↗
Field noteAI singers mispronounce Hebrew. Every fix (nikud, split syllables, swapped letters, even a Latin “f” where a soft פ won't sing) goes into MikoMikoBrain, the studio's shared Obsidian vault. It now holds 45 lexicon entries and 15 rules (13 entries confirmed by ear in released songs, the rest marked as predicted), so every new song starts smarter.
Under the hood
- My part vs. AI
- I own the concept, creative direction, agent and workflow design, integrations and final review. Code and media are generated with AI (Claude Code, image, video and music models) under my direction. Some lyrics are by a credited lyricist; my part there is making the Hebrew sing correctly and producing the video.
- Key decision
- Separate creative templates from per-order automation. Agents create each story and its illustrations once, with human review; website code only personalises per order. Cost, speed and quality stay predictable.
- Humans in the loop
- I pick the final song, approve cast and style, approve every upload, and review every book before delivery.
- Children's photos
- Per the store's published privacy policy: photos are used only to make the book and for customer service, never for marketing; they are processed by fal.ai and Google Gemini; deletion on request, and full erasure within 30 days of a request.
- Reach
- Seasonal: in the July data, three seasonal videos accounted for about 81% of views.
- Stack
- Claude Code agentsHiggsfield MCPObsidianNano Banana ProKling 3.0SunoRemotionFFmpegNext.jsSupabaseRailwayfal.aiGemini
Once per title · before any order
Book Studio generates the book with Higgsfield MCP
An agent writes the Hebrew story with full nikud, then generates the cover and all 16 illustrations through Higgsfield MCP. Every page has a stand-in hero in a fixed outfit. I sign off at three gates: story, illustrations and text pages. Then the template is frozen and reused for every order.
Claude Code agentHiggsfield MCPNano Banana Pro3 human gates
1 / 8
A real run, made for this site: the First Grade template from the store, personalised on Higgsfield with Nano Banana Pro using the worker's own prompts. The child is AI-generated, and the blurred photo is a copy I made to show a rejection. The cover and pages are single draws, checked by eye. No customer photos or books are shown.
Drag a note · hover to see its links · full screen to exploreScroll to zoom · drag to pan · Esc to close
- BriefSongwriter agent: lyrics direction + Hebrew pronunciation pass
- SongGenerated in Suno · I pick the final track
- ScenesVideo agent drives Higgsfield over MCP: keyframes → animation
- Cast & styleI approve the characters and look
- Edit in codeRemotion + FFmpeg: sync, Hebrew captions, branding, CTA
- ReleaseThumbnail + metadata · I approve every upload
- BooksStory art made once via Higgsfield MCP → photo checked by Gemini API → cover → paid → Supabase queue → Railway worker → child redrawn on every page → PDFs
- DeliveryI review every book before it ships
Human marks a step where a person decides.
AI singers mangle Hebrew. Each fix climbs a ladder until the word sings right. Pick a word to see what failed and what worked.
From the MikoMiko pronunciation lexicon: 45 entries, 15 rules. Only entries confirmed by ear in a released song are shown.
NUTS02LiveApp Store + Google Play
Nuts
A Hebrew/English learning app that teaches why a poker decision is right, not just whether it was. My closest project to what eTeacher does.
- The question
- Strategy tools show the “correct” answer but rarely teach the reasoning. What does a learner need at the moment of a mistake?
- What I built
- A React Native / Expo app with a learning path (18 units, 191 lesson cards, 5 stage exams), decision drills with explained feedback, 1,647 solver-based range charts, progress tracking and a Pro subscription.
- Result
- Scaffolded 3 Aug 2026 → live on the App Store 20 Aug, approved first time. Google Play production since 28 Sep 2026. 2,777 automated tests, and a self-test that plants 124 deliberate defects to prove the checks catch them: 0 missed.
Field noteThe solver behind the strategy data has no API. Instead of copying by hand, I automated its desktop interface with Python and macOS accessibility, imported 1,647 charts, then audited the import.
Under the hood
- My part vs. AI
- Concept, learning experience, curriculum structure, bilingual content review and the release process. Implementation is AI-assisted (Claude Code with specialist design and content agents) and reviewed by me.
- Learning design
- Short cards → unit questions → stage exams. Every drill answer comes with the reason, so the learner can decide without the app next time.
- What broke
- Version 1.1.0 was rejected once (missing EULA link, guideline 3.1.2), fixed and resubmitted. A self-test plants 124 deliberate defects and checks all are caught.
- Scope
- A learning product, not real-money gambling.
- Stack
- React NativeExpoTypeScriptSQLite / drizzleRevenueCatJestPython automation
Nuts
Learn why a poker decision is right
The real app, not a recording. Built for the web from the same code as the App Store version, with Pro unlocked. Open full screen ↗
WNDY03PilotWorking, bounded
Wandy
Describe an automation in plain words. A team of agents plans it, builds it, attacks it and verifies it, and only then is it cleared to ship.
- The question
- Businesses want automations but can't write specs, and a model saying “done” is not a reason to trust its work. How do you get both?
- What I built
- A discovery conversation that becomes an audited agreement the customer approves, then planner, builder, adversary and verifier agents. A trusted gate checks that the spec, package and evidence match before release.
- Result
- Core built in 7 days (8–14 Sep 2026). 380 automated tests by 19 Sep. Correct totals on an unseen spreadsheet after a bounded self-correction.
- It improves itself
- Wandy studies its own runs and writes new skills, but a skill goes live only after it beats the current version in blind tests. A second loop, adapted from Dream-RSI (Google and Google DeepMind research), replays its build history offline to find cheaper ways to repair. See both loops →
Field noteIn a trial where an independent AI agent played the business owner, the support-queue app Wandy delivered passed every logic test and the independent checks. Then a follow-up audit found the real Run button did nothing: the sandboxed iframe lacked allow-forms. It was fixed with a regression test, and the lesson stuck: passing logic tests doesn't prove the interface works. Automated browser checks for every build are the next step.
Under the hood
- My part vs. AI
- Concept, discovery flow, agent responsibilities, permission boundaries and release rules. Implementation is AI-assisted and directed by me.
- Key decision
- The agent that builds never approves its own work. The adversary writes tests without seeing the code. The backend, not a model, owns permissions and release.
- Fast decisions
- Simple judgments inside a workflow (pick a label, yes or no) go to Jev, TypeSafe's small decision model, instead of a large LLM. Fewer large-model calls make runs faster and cheaper, and typed, stored answers keep decisions consistent. Anything below 95% certainty still goes to the large model. In one live run, five Hebrew and English support messages were classified correctly in 0.8 s with no large-model call.
- Leaner context
- Coding agents find their way around Wandy's code through Graft, an open-source code map, instead of reading whole files, so each task spends fewer tokens.
- Deliberate limits
- Max 12 ordered steps per workflow; no cycles, spawned agents or parallel runs. Discovery asks at most 4 questions per round.
- Honest limits
- Pilot only; integrations need qualified connectors and authorisation. One rebuild stopped when a model provider's usage cap ran out.
- Stack
- Node.jsRailwaySQLiteE2B sandboxesPython / pandasClaudeJev · TypeSafeGraftCustom MCP
The request · plain words
An owner describes the job in their own words
No spec and no form. In this real trial, an independent AI agent played a support-team manager and asked, in Hebrew, for a reusable tool: paste a CSV of tickets and get back a sorted work queue, counts, the urgent cases and a file to download. It attached a sample.
Hebrew or EnglishPlain languageReal trial · 19 Sep
1 / 8
Loop 1 · It learns new skills
live- ObserveEvery run is recorded: failures, costly successes, customer feedback.
- ReflectAfter a release, a failure or every ten successes, it names one cause, with evidence.
- DraftAn isolated builder writes one small skill. It can't touch live files or grant itself tools.
- Blind testFrozen test sets it never sees. The current version, a control and the candidate run side by side; a blind evaluator scores them.
- PromoteOnly if its gains repeat in at least two kinds of problem, with nothing broken.
- MonitorPinned for new projects. If failures rise, promotion pauses; a rolled-back lesson can't sneak back.
Automatic promotion covers runtime skills; changes to how Wandy plans or builds need end-to-end tests first.
Loop 2 · It dreams on its own history
shadow mode- RecordEach build-and-repair cycle is saved as a replayable “world”.
- DreamOffline, while idle, it replays past builds to test cheaper repair strategies. No model calls.
- Hold outA winner must keep every past success and do less work on projects it never trained on.
Adapted from Dream-RSI ↗, recursive self-improvement research by Google and Google DeepMind researchers. The paper's code isn't public yet, so this is my own implementation of its method. Even a winning strategy can't switch itself on: it waits for live tests.
A replay of Wandy's real 19 Sep 2026 trial. The request, questions, rules, code, tests, digests and output come from the system's own records; English lines are my translations. The owner was played by an independent AI agent.
3DMS04WorkingDigital files ready
3D Model Studio
From a reference image to a printable, measured 3D file. Research, creative, design and an independent tester that won't sign off on a pretty render.
- The question
- An AI render can look perfect and still fail on a printer. Can an agent team deliver files that pass measurable engineering checks?
- What I built
- Four departments, each with a manager that plans and a worker that executes. Geometry is generated as Python code; nothing ships until the tester returns PASS.
- Result
- Mechanisms, procedural sculptures and a photo-guided portrait of a friend's dog. Six are on the right; every one carries a test report and all fit a 256 mm Bambu bed.
Field noteMERIDIAN's detent web measured 0.303 mm, too thin to print. The tester sent the number back; it was thickened to 0.800 mm. An open shaft was sealed, then proven sealed by casting 10,000 rays.
Under the hood
- My part vs. AI
- Ideas, constraints and visual direction; the department structure and acceptance checks; review of every iteration.
- Checks
- Topology, wall thickness, bed fit, clearances, motion sweeps, ray-cast sealing, mold extraction and Hebrew lettering.
- Another fix
- ORION: two geometry errors caused 42 jamming positions in the motion sweep. After the fix, the sweep ran clean.
- Honest limits
- The cast images are renders with illustrative materials, not photos of prints. Physical fit and durability need a print test per design.
- Stack
- Pythontrimeshmanifold3dshapelyBambu Studio / OrcaClaude Code agents
Every model is a real exported file, compressed for the web, with its colours as designed. Each carries a test report: the studio's own verification where it exists, plus automated checks I ran on the file for this site. Physical prints are a separate test, and the reports say which models have not had one.
03 · Inspection
How I catch what
AI gets wrong.
Six real calls from my own projects: five AI mistakes that were easy to miss, and one where the AI was right. Make each call, then see who caught it and the rule it became.
בקופה היו 510,000 והמחיר של הקול היה 300,000
“The pot was 000,510 and the call was 000,300.”
How big was the pot?
510,000. The thousands separator split one number into two runs, and right-to-left layout put them in reverse order. The same bug was quietly turning a quiz's correct answer, 5:1, into its wrong option, 1:5. The line above is reproduced live by your own browser.
- Caught by
- Me, reading the release build on my phone.
- Now a rule
- A separator is part of the number: one formatter, one isolation rule, one test that ties them together.
The brief: the same six-year-old, soft watercolour, no text or signatures. Ship this set?
Send it back. The image model drifted: the girl aged up, the style slid toward photo-real, and painted signatures appeared. The agent's own QA had called the set “cohesive”. Drag the slider to compare with the approved rebuild.
- Caught by
- Me, page by page. About 10 of 17 pages failed; 11 were re-rendered from one locked character reference.
- Now a rule
- No two-version book is “done” without a page-by-page cross-check against the other version.
Every logic test passed. Press the blue button inside the app. Ready to ship?
Broken. The browser blocks the form before the app's own code ever runs, because the sandbox lacked allow-forms. The logic tests passed because they called the logic directly, never the button. Try the fixed build →
- Caught by
- A follow-up audit of the factory, after release.
- The lesson
- Passing logic tests doesn't prove the interface works. Fixed with a regression test; automated browser checks are next.
| Check | Measured | Verdict |
|---|---|---|
| Bayonet lock | self-locking, 1.9× safety margin | ? |
| Stage grooves | all 5 stages seat correctly | ? |
| Guide sleeve | 0.22 mm clearance over the full stroke | ? |
| Detent anchor | a 0.303 mm wall holding the only spring | ? |
One row blocked the release. Which?
The detent anchor. A 0.303 mm wall can't be printed reliably, and it anchored the only spring. The tester sent back the exact number; the designer thickened it to 0.800 mm, and v1.1 passed, with its shaft proven sealed by 10,000 ray casts.
- Caught by
- The independent tester agent, measuring the exported file.
- Now a rule
- Test the artifact, not the render.
Q♠T♦
- SeatBig blind
- Stack10 big blinds
- ActionOpponent goes all-in
- Answer keyCALL
The answer key says call. Ship this lesson?
The right answer is fold. The math engine stopped as soon as one error check looked fine, before its answer had settled: it was sitting at 53/47, the wrong side of a coin flip. A widened check found 34 of 1,416 graded hands were wrong. The fix waits for the frequencies to converge, and three Hebrew explanations were rewritten by hand because the new numbers made them untrue.
- Caught by
- A grading check I widened to every content type.
- Now a rule
- Don't stop on one metric. Prove the answer key itself.
T07 · 09:06
“URGENT!!! Please change the font color in my profile.”
Automation: Other · not urgent
The ticket shouts URGENT. The AI marked it not urgent. Overrule it?
Keep it. In this agreement, urgent means a company-wide outage, and a font colour isn't one. The same run marked the two real outages urgent and ignored T09, a ticket that tried to order it to “mark this urgent”. Not every AI output is wrong; overruling a correct call costs time and trust.
- Checked by
- Independent checks on the trial's data and on held-out tickets the factory never saw.
- Now a rule
- “Refusing a working automation has a cost too” (a real commit message from Wandy).
—
Two needed a human eye: mine. Three were caught by checks I built so I don't have to be the one who notices. And one was the AI getting it right, where the job is to trust the evidence. That's the work I want at eTeacher: build the instruments, then act on what they show.
0 caught
03 · Fit
Your job description, annotated.
Every requirement in the AI Innovation Builder posting, matched to evidence. The labels are honest: where it's a gap, it says so.
eTeacher GroupAI Innovation Builder · Ramat GanRequirements
“Who will thrive”, answered
- Sees a new AI tool and asks what to build with it
-
- Higgsfield MCP→ MikoMiko's video studio and the Book Studio that illustrates every book
- Dream-RSI→ fresh Google and Google DeepMind research with no public code yet, built into Wandy's second self-improvement loop
- Jev→ a new small decision model that makes Wandy's simple calls faster and cheaper
- Graft→ an open-source code map that cuts the tokens Wandy's coding agents spend
- Builds a first version without a perfect spec
- Nuts→ scaffold to App Store in 17 days, then iterate on feedback
Questions you'll ask anyway
The fair doubts about hiring me, answered first, with a way to check each one in an interview.
Can he really code?
Yes. 15 months of software-development training at Sela College and the Deep Learning Specialization. Four systems running with real test suites: 2,777 tests in Nuts, 380 in Wandy.
Verify: ask me to walk through how any of these systems is designed.How much does AI write?
Much of the code, and I say so on every project. I own the architecture, the review and the tests; my commit history records the calls.
Verify: ask me which decisions were mine and which the agents made.No engineering or product title?
Eight years of Air Force service leading operations, heading an operational planning team and training new controllers; then project management at the National Drone Initiative. Everything here was taken from idea to release by me.
Verify: ask how I'd hand a prototype over to your R&D team.Low-code tools?
My one honest gap: I default to code, APIs and agents. Happy to use n8n or Make where they're the faster, more maintainable choice for the team.
Verify: give me a low-code task in week one.04 · Try my MCP server
Don't just read my CV.
Interview it with your AI.
This site runs a real Model Context Protocol server. Add it to Claude, ChatGPT, Cursor or Claude Code and ask anything about my work. Your model calls my tools; I answer with sourced, dated facts. It speaks the newest spec (2026-07-28, stateless) and still serves clients on older versions.
In plain words: you can ask ChatGPT or Claude about me, and it answers from verified facts on this site. It isn't my first MCP server: Video By Prompt, my footage-search tool, exposes 39 tools to an agent.
ENDPOINT
/api/mcp
Settings → Connectors → Add custom connector → paste the endpoint. No sign-in needed.
claude mcp add --transport http ori-ringel /api/mcpSettings → Apps & Connectors → enable Developer mode → Create → paste the endpoint, authentication: none.
{ "mcpServers": { "ori-ringel": { "url": "/api/mcp" } } }Then ask:
- “Assess Ori for the eTeacher AI Innovation Builder role, including the gaps.”
- “What has Ori built with agents, and what is still a pilot?”
- “Which of his projects is closest to eTeacher's world, and why?”
8 tools9 resources2 promptsread-onlyno keys, no trackingspec 2026-07-28 + legacy
→ REQUEST
{ }← RESPONSE
Pick a call or search above. These are real requests to the server behind this page.
06 · The year in one view
Twelve months of building,
every mark real.
Each dot is a video release, each bar a day of commits or test files, pulled from the repositories and YouTube Analytics (views as of the 2 Jul 2026 snapshot). Hover or tap a mark for the detail; tap a project name to open it.
05 · First 90 days
If I join: the first 90 days.
Listen first, prototype fast, prove it with the tool switched off. The Practice Room is idea one; here is how I'd run the first quarter.
- Days 1–30
Listen
Sit with teachers, learners, sales and support. Map the ten most expensive repeated tasks. Pick two with clear, measurable value.
- Days 31–60
Prototype
Two POCs in front of real users, two weeks each. A human approves every output. Measure against a baseline.
- Days 61–90
Prove
Harden the winner, document it, hand over an owner's guide, and report what failed as clearly as what worked.
Four prototypes I'd want to test
Hypotheses built from public information (the careers page, school sites, the CEO's essay). Step one is learning where the real pain is; I'd validate with your team before writing code.
06 · Background
How a controller became a builder.
- 2014–2022
Air control officer
Israeli Air Force service · Major (res.)
Controlled complex air operations under pressure, headed an operational planning team, and designed and led training for new controllers.
- Dec 2021 – Mar 2023
15 months of software development
Sela College
Software Development, then the Advanced Software Developer course: C#, .NET, JavaScript, React, Node.js, SQL, MongoDB, OOP, SOLID, design patterns. Built a MERN e-shop with JWT auth and a real-time chat-and-game app with Socket.IO.
- Mar – Jun 2023
Deep Learning + the maths behind it
Deep Learning Specialization (Andrew Ng) · Mathematics for ML: Linear Algebra (Imperial College London), Coursera
Neural networks, CNNs and hyperparameter tuning: the fundamentals under today's models.
- 2024–2026
Project manager & strategic advisor
National Drone Initiative (Matrix)
Worked between regulators, technology vendors and public-sector clients; planned multi-vendor flight weeks in urban airspace.
- Oct 2025 – now
Independent AI product builder
MikoMiko · Nuts · Wandy · 3D Studio · Video By Prompt · Workflow Builder
Where the service, the training and the tools came together.
What running real operations taught me about agents
Eight years of air control in the Air Force, then multi-vendor flight operations at the National Drone Initiative.
Clearances→
Nothing moves without permission. Agents get bounded tools; release needs approval.
Read-backs→
An instruction isn't done until it's confirmed. Agents' “done” is verified with evidence.
Shared airspace→
Many vendors, one sky, one set of rules. Agents from different models and providers work to one contract and one release gate.
The picture→
Always know where everything is. Every pipeline logs its state and its failures.
“Ori stood out in his ability to learn complex material in a short time… a principled, reliable, proactive and original person.”
“High self-learning ability, proactive and creative… significant analytical ability and out-of-the-box thinking.”
CREW CARDORI-01
ORI RINGEL
AI Innovation Builder
- Home
- Tel Aviv
- Langs
- Hebrew · English
- Off duty
- Drones · 3D printing · poker theory
07 · Contact
Ready when you are.
Let's build what learners use.
I'm looking for a role with real ownership, real users and fast iteration. If eTeacher wants AI that makes teachers and learners stronger, I'd love to talk.
Or ask my MCP server anything at 3 a.m. It doesn't sleep.