Budō AI: A Real 3D Sensei, Five Agent Masters, and the Form You Can Finally See

By the SynaptixLabs Team

---

Open a browser tab and a karate master is mid-kata in front of you.

Not a cartoon. Not a stock animation. A real 3D human, reconstructed from an actual Heian Shodan video by Meta’s SAM 3D Body, standing in a dim dojo and moving through the form. Drag to orbit him. Strip the gi back to a gold wireframe. Pull all the way through to the 127-joint skeleton underneath. This is the dojo, rebuilt for the screen — for the long, quiet hours between two weekly classes when there is no one to correct you.

That gap is the whole problem, and it is worth naming precisely.

The eye you cannot get at home

Great martial-arts training needs an expert eye on every stance, every hip rotation, every guard line. A good sensei watches you move and says the one thing that unlocks the next month of practice: your front stance is too shallow; rotate the hip, not the shoulder.

A teenager practicing alone at home, between two weekly dojo sessions, cannot get that eye. So the kata drifts. A stance is misremembered, a sequence is half-forgotten, a bad habit sets like concrete — and the student plateaus, or quits, because solo practice feels unproductive and unseen.

The tools that exist do not close this gap. Generic video-analysis apps have no martial-arts knowledge — they can count reps, not judge a kokutsu-dachi. Karate apps have crude 3D and no coaching. And in Hebrew, the entire category is simply empty. No product combines a structured curriculum, demonstration-quality 3D, contextual coaching, and video comparison — in any language, let alone for a Hebrew-first teenager in Tel Aviv.

Budō AI (武道) is built to fill exactly that hole: a bilingual, Hebrew-first (RTL) and English AI martial-arts dojo, launching in Israel for teenage Shotokan students aged 12–17, with Heian Shodan as the opening wedge.

First, make the form visible

You cannot coach what the student cannot see. So the foundation of Budō AI is demonstration: a 3D master good enough that an instructor would certify it.

This part is real today, and it is the most honest thing we can show you. We took a real JKA Heian Shodan demonstration on YouTube — 536 frames at 2 frames per second — and ran it through Meta’s SAM 3D Body, which lifts a single image into a 3D human. The output is a 127-joint MHR mesh, hands and face included, exported as a roughly 18,439-vertex glTF avatar. Dressed in a gi, planted on a dojo floor, lit like a bronze studio statue, it performs the kata live in WebGL.

The studio below is that avatar. It is a self-contained interactive kit — a separate Three.js build made for the open web (the production app renders its 3D in Babylon.js; this is its own thing). Drag to orbit, zoom, change speed. Switch render modes: Gi, Wireframe, Vertices, Skeleton, and Compare. Skeleton mode peels the avatar back to the 127-joint rig the coaching reasons over. Compare overlays a synthetic student on the sensei’s skeleton and colours each joint by how far it drifts — green on target, gold for a minor miss, vermillion for fix this — then lets Yuki, the compare sensei, score the form.

Read the honesty bar under the viewer before you read the scores: the avatar is real SAM 3D Body output; Compare uses illustrative demo data — the live app scores your own video. That distinction is the point of the rest of this story.

Figure: the Budō AI “DAN 10 Studio” — a real SAM 3D Body avatar of Heian Shodan, with five render modes and Yuki’s Compare scorecard. The avatar and its 127-joint skeleton are real reconstruction output; the Compare student, its deviations, and the scores are an illustrative synthetic loop (re-rollable) — in the product, you upload your own video. The five sensei seals and the SAM 3D pipeline panel are honest depictions of the architecture described below.

Five masters, one contract

Here is where Budō AI stops being a clever 3D viewer and becomes a thesis about how to build software.

The coaching is delivered by five named senseis, each a specialist: Takeshi (武, the guide), Kenji (剣, kata technique), Yuki (雪, compare), Haruki (春, history and culture), and Rin (凛, drills). It is tempting to imagine five chatbots. That is not what they are.

The five senseis are five typed Agent Types sitting behind a single Agents-as-a-Service contract. Budō’s own backend — a Python 3.12 FastAPI service — never calls a language model directly. That is not a style preference; it is the project’s first architecture non-negotiable: all AI goes through the AGENTS platform, with no direct LLM calls in the app. When a student needs coaching, the backend does not reach for a model. It ships context — your locale (he or en), which kata, where you are in the app, which step — across one governed seam to the platform (named AGENTS in the repository; productized on our site as Nexus), via a single REST call: POST /api/v1/execute.

Two things then happen on the far side of that seam, and both are quietly radical.

First, the platform picks the master — not the student. This is governance by routing. A student in the kata studio is routed to Kenji; a student in Compare gets Yuki; a newcomer in onboarding meets Takeshi. Nobody opens a model picker. Nobody chooses a personality. The right specialist arrives because of where you are, and the senseis’ prompts live as versioned Markdown in PostgreSQL — editable, reviewable, and shippable without a redeploy.

Second, the platform picks the model. Behind the same contract sits a Model Router (Gemini primary, Claude secondary). The provider can change without Budō changing a line, because Budō consumes a contract, not a vendor.

The deepest part is that the same harness powers two very different features. Conversational coaching and Compare-mode video scoring are not two systems — they are one. In the MVP target, Compare sends your uploaded kata through the backend to Gemini Vision and back to Yuki, who returns the three things to fix, in degrees off ideal, with drills. The studio above shows that loop with synthetic data so you can feel its shape; the engine behind it is the same single seam the conversation uses.

Consume the platform; do not rebuild it

Why route everything through one seam instead of wiring a model straight into the app? Because that is the entire reason a platform exists.

A product team that embeds its own model inherits the hardest, fastest-moving, most expensive problems in software — inference cost, provider churn, prompt governance, safety, evals — and inherits them again for every new feature. Budō AI refuses that. It owns the dojo: the curriculum, the belts and runes, the 3D, the bilingual RTL experience. It consumes the intelligence. The masters carry their hedging rules with them across the seam — “it looks like your stance is…”, never a false certainty — and one rule above all: a sensei never claims to replace your live instructor. Budō AI is for the hours between classes, not instead of them.

This is also why a single new product can prove a platform. An explicit goal in Budō’s plan is to “prove AGENTS-as-a-Service in a new vertical.” Martial arts in Hebrew is about as far from the platform’s other tenants as you can get. If five typed senseis can context-route, model-route, and stay governed here — with their prompts hot-editable and their model swappable — then the seam holds anywhere.

Where Budō AI actually is

Honesty is a non-negotiable here, so let us be exact about status.

Budō AI is in development — and further along than that phrase usually implies. Sprints 00, 01, 02a, and 02b are closed; Sprint 03 is active. Sprint 02a shipped a working application on GCP staging in Tel Aviv (Cloud Run in me-west1, Cloud SQL Postgres, Cloud Storage, Firebase Auth) with authentication, a dashboard, the belt and rune progression, all five masters, and 931 tests passing — 388 pytest, 502 vitest, and 41 Playwright. The architecture is the three-tier one this article describes: a Next.js 14 frontend, the FastAPI backend, and the AGENTS platform consumed as a service.

Two caveats matter, and we will not bury them.

The senseis are built to run as governed agents on Nexus — that is the target and the architecture. Today, the backend’s masters module ships against an AgentsMockClient with canned responses (“no external AGENTS dependency in Sprint 01 — mock client only”), with direct Vertex/Gemini documented as the fallback if the platform is not ready when live wiring lands. The seam is real and load-bearing; what flows through it in production is the next step, not today’s.

And Compare — per-student video scoring — is the MVP target (tracked as R-BD-012/013), not a shipped feature. The studio above is honest about this: a proof of concept, where the avatar is genuine reconstruction output and the Compare scorecard is an illustrative synthetic loop. Compare mode ◍ — live pose feedback coming soon. Because the product serves minors and is Hebrew-first, it is held to real privacy and accessibility standards, and there is no real student data, video, or name anywhere in this demo by design.

The form you can finally see

Strip it all back and the shape is simple. A teenager will always need a human sensei. What they have never had is an expert eye in the room at home, on a Tuesday night, when the dojo is dark and the kata is drifting.

Budō AI builds that eye in two moves: make the ideal form visible with a real 3D master anyone can orbit, and make the coaching governed by routing every word through one Agents-as-a-Service seam — so the right sensei and the right model arrive by context, the prompts stay editable, and the product never has to become an AI lab to feel like one.

Perfect the form. 型を極める.

---

Budō AI is a SynaptixLabs product, in development. Its senseis are built to run on Nexus — our Agents-as-a-Service layer — because a product should consume intelligence, not reimplement it. Talk to us.

— The SynaptixLabs Team