Turn Lab
Pick a clip that already has a saved analysis, or add a new one. Snowboard clips only — the lab reads board edge from pose.
A new clip runs a full review first, because the lab reads the pose that review produces. It lands in the list when it finishes.
How this report was made · admin only
🧪 Sandbox draft
Loading…
🎯 Pipeline designer — flowchart of what a review runs
Each realm is a flowchart: 🎬 video flows down through blocks to 📄 report. Drag a block from the palette onto the canvas to add it; drag a node to move it; drag from a node's bottom ● to another node to wire an arrow. ◇ Add decision drops a branch diamond; label its arrows Yes/No (drawn + saved now — the review follows branches in the next update). Click a block to edit its prompt right here; double-click a diamond to set its question. Select a node/arrow, then click Remove selected or press Delete. Use ⧉ Duplicate to make an editable copy of the current realm.
🧪 Prompt Lab agents — the judge & the tuner
The two AI agents behind Stability / Batch runs, editable right here like any prompt. The verdict judge decides stable / wobbly / unstable; the tuner reads that verdict and proposes prompt fixes. Machine parts (injection safety + output format) stay in code, so you can't break them. Blank + Save = back to default.
Gemini is the default and the only one tested here. Add other companies (Qwen, DeepSeek, or OpenRouter fronting Claude/GPT) in Settings → Add a provider, then pick them here.
Prompts are grouped by sport, then by pipeline step. Drag a prompt to reorder it within its step — the order is the sequence the AI reads them, and it changes the result.
📝 Run Logger — field notes from the phones
Every cloud ride for the chosen day, with its note: tags, written text, voice memos, and the laps the rider marked mid-ride (each lap can carry its own tags, text, and voice). Highlighted tags are one-off entries the rider typed — future catalog candidates.
📖 Terminology library
One agreed meaning per manoeuvre, movement, condition and feature. Write a term's name in any prompt — "carving", "sliding" — and both its definition and its recognition rule are attached to that AI call automatically, so you never re-explain it. Different words for the same thing (skidding / sliding / smearing / 搓雪) all resolve to one term, which is what stops the coaching drifting.
Only the terms a prompt actually uses are sent — but they are not free. Each one carries its full recognition rule, so a prompt that touches many terms can add more text than the prompt itself. The block is capped at 12,000 characters per call and anything over the cap is named but not spelled out. Keep a term switched off unless you want the AI to be taught it.
+ Add a new term
API keys
Paste a key, it is checked with the provider, then stored in the engine's .env on this Mac.
Keys are write-only — they are never shown again. This admin page only works on the Mac running the engine.
AI engine
Gemini watches the whole clip — it does not just look at a few stills. The two knobs below do different jobs: the active one for your provider is bright, the other is dimmed.
Frames cost money. The video part of a review is billed per frame, so this dial multiplies the cost of every review — 24 fps costs about 12× what 2 fps costs on the same clip (roughly 71 input tokens per second of clip, per fps). Higher may read fast movement like edge changes more closely, but that is unproven — use Sandbox → Stability on one clip at two settings before leaving it high.
Gemini (default): watches the full video at the rate on the left, then picks the coaching moments itself. The number of screenshots in the report is set by “Coaching screenshots” below, not here. Screenshot-only AIs (Claude, Qwen…): cannot watch video, so they only ever see the stills set on the right.
Add a provider
Any service that speaks the standard OpenAI-compatible API works: Qwen (DashScope), Doubao, DeepSeek, Kimi, OpenRouter, a local model… Enter the base URL (the part before /chat/completions). Added providers review sampled screenshots — only Gemini watches the whole video.
Report format (what the user receives)
Library storage
Off (default): the Library keeps the compressed video — cheap (~$0.02/mo per 100 reviews) and still fully watchable. On: keeps the original full-res upload.
Every analysis is saved here automatically. Sort and filter to find a past review.
Improvement
Turn administrator grading into measured prompt improvements. Automation may prepare a candidate in Sandbox; an administrator starts held-out evaluation and decides whether it moves into production.
Automation stops after preparing a Sandbox candidate. An administrator starts evaluation, and production promotion always needs explicit confirmation.
Review Queue
Finish grading AI output that still has unanswered sections. Judge whether the AI is accurate—not whether the rider's movement is good or bad.
Loading…
Insights
Patterns are grouped by the prompt, model, realm, and reason behind the human grades.
Loading…
Experiments
A candidate must be tested against held-out reviews before it can become approval-ready. Testing never edits production.
Loading…
Controls
A failed regression, missing evidence, or exhausted cost limit blocks promotion. “Cannot judge” grades improve coverage tracking but never count as correct or incorrect.
Standard test set — the prompt regression loop
Flag a few pool clips as ★ standard, edit the Sandbox, then Generate to run the draft across all of them before promotion. Every run is saved to the Library and clearly stamped as Sandbox.
Compare runs each ★ clip on the Sandbox draft and pairs it with that clip's last saved run from the Library — so the old prompts aren't re-run (half the cost). Changes are highlighted.
Prompt Lab — is the prompt stable, and does it generalize?
🎲 Stability runs one clip several times on the same Sandbox draft and checks the answers agree. ▶ Batch runs different clips on that same draft. Production stays untouched until promotion.
The judge and tuner agents behind these runs are edited on the Sandbox tab, next to the pipeline designer.
Add a clip to the pool
Upload a clip once, re-analyze it any time without re-uploading. Stored durably on the engine.
Gold examples — teach the rubric
Build your evaluation form with any field types, then add videos with the expert answers filled in. Later the AI learns to reproduce these reports — the goal is ≥95% closeness to your gold.
Runs the AI on each gold video, fills your form, and scores how close it gets to your answers — field by field. Goal: ≥95%. ~2 AI calls per example (only when you click).