unvaultlabs
Checking endpoint…Shows whether the Unvault model inference endpoint is reachable and serving. The indicator reads a cached snapshot — it does NOT probe HF on every poll (that would prevent the endpoint from scaling to zero and burn budget). States: checking (initial probe), offline (unreachable), presumed cold (no activity in 15+ min, or model not listed), loading / warming (~30–90 s cold-start), and warm (model serving — confirmed by a real request). Send any inference request to refresh.
How Labs Works

Tools that make our model better, every day.

Unvault Labs is the internal workshop for our jewelry AI — inspect pieces, search the forge, annotate the hardest cases, and watch the model get sharper. Every correction you make is shipped back into the next training run.

Sign in to startView model status
⚔️Tournament Complete

THE GRAND TILT

Eleven Houses · Seventeen Rings · One Champion

Gemini38.0
Fable34.0
Grok34.0
Champion: Gemini→
The Grand Tilt — AI jousting tournament broadside
Comp Shopping

See what designers it compares to.

Browse comparable pieces from the forge and spot the brand premium at a glance.

54,381 comps in the forge right now.

Browse comps→
The Tools

Seven workshops. One sharper model.

Visual Inspection

LoupeUpload a jewelry photo to inspect hallmark, metal, weight, and category predictions from the vision model.

Drop a photo. Get weight, metal, purity, category, and confidence — then correct what we got wrong.

Open
Semantic Search

SearchDescribe a piece in plain English and find visually similar items across the forge database.

Type a designer, a style, a feeling. Find visually similar pieces across the entire forge.

Open
RLHF Annotation

GymReview the model's worst misses, vote on corrections, and send feedback into the next training run.

The training floor. Work the queue, correct the model, watch your impact compound over time.

Open
Deploy Status

ModelSee which model version is live, check accuracy, and track the next deployment.

What version is live, how accurate is it, what's the next training run. The pulse of the system.

Open
Single-Item Appraisal

AppraiseGet a structured, streaming appraisal for one item — weight, metal, category, brand, and confidence.

Drop a photo (or jump in from any Search result). Get a structured appraisal — weight, metal, category, brand, confidence — streamed in token by token.

Open
Small-Batch Evaluation

EvalRun a small batch of test items, then view MAE and per-field error tables to validate a new build.

Drop a CSV of test items, run the lot, get MAE, accuracy by field, and a per-row diff table. The workhorse for validating a new build.

Open
Prompt Lab

Face-offPit two prompts against each other on the same input and build a win-rate leaderboard.

Two prompts, same input, blinded. Pick a winner, build up a win-rate tally. The fastest way to iterate on system prompts.

Open
Recent Activity

What's moving today

View model status →

No recent model runs recorded yet.

Activity will appear here as new training runs and deployments land in unvault.model_runs.