Skip to content

Demo walkthrough

Sixty seconds through what is built, in the order a person meets it, on a stack started with ./dev up in fixture mode (no key: every model answer is a recorded one and nothing is spent). With a key, the same steps run live. The demo set is a drawer of idle things, photographed by the owner: three monitors, two ThinkPads and a spare laptop board, an old graphics card, a slot 1 processor, a powerline adapter still in its wrap, and the worn agitator of a frappé maker.

The set, loaded

demo/set.yaml names the things in the walkthrough: each one's photos, what the owner knows about it, whether a draft or price guidance is written for it, the searches with the listings brought back (written examples labelled as such), and the proposal brief. Three commands:

./dev demo-scrub ~/raw-photos   # re-encoded copies into demo/photos/, WAV notes into demo/notes/
./dev record-fixtures           # with a key: the inventory reset, one run, every answer kept,
                                # then every reading a sample click can ask for (each photo
                                # alone, each item's photos together, with the label found first
                                # and without); a run that stopped half-way picks up where it was
                                # (what is recorded is replayed, not bought again) and fails when
                                # a part of the set was skipped; delete a fixture file to record
                                # it afresh
./dev demo                      # the inventory reset and the set loaded, in fixture mode

./dev demo leaves the stack with the whole set in the inventory, every observed value accepted, drafts, guidance, searches with their candidates and the proposals in place, so every take and every visitor starts from the same screen. Once the fixtures for the set are recorded, the demo spends nothing: the service line at the foot of every page says recorded answers only (nothing spent). The fixtures are keyed on what is sent, so a change to EXTRACT_ENHANCE or IMAGE_MAX_PX means recording again. On a gated host the public demo is the set, not an open reader: a visitor without the code picks the set's photos in a gallery instead of uploading (see Demo access), and each answer is the recording of that exact photo; a selection with none says so. docker compose run --rm backend python -m app.demo samples --set /demo/set.yaml checks in fixture mode that every sample has its recordings. The set loads with whatever files are present; a photo named but missing is skipped with a note.

An item may name a picture: (a template id such as smart_mirror, or a few words), drawn by Muse Image after its drafts; four items of the set do. A drawing is a paid call, so a live load draws it only with PICTURE_ENABLED=1 (./dev record-fixtures sets it for its run). Fixture mode shows a drawing only once a record run has kept it, and until then the load says not recorded yet for it and shows none, never the stub image as a drawing.

1. Read a label and save it

Read a label takes up to four photos (signed out on a gated host, a gallery of the set's photos to click instead, one reading per item or photo). With Find the label in the photo first on, SAM 3.1 looks for the item, the label and the connector: the overlay shows the regions it found, with a note for each it did not, and when no label is found the photos are read whole. Muse Spark reads the label crop, or the photos, into typed fields, each with its confidence and its source. Save privately makes it an item in My inventory: every observed value waits for you to confirm or reject it, and anything you add beside it carries your word as its provenance.

2. Measure the thing that decides fit

On the agitator's page, Measure a feature: the rod photographed from the side with a ruler along it, two marks on the ruler and two on the rod, and the number comes from geometry, with its uncertainty and the method. A blurred or oblique capture is refused with the reason. Measure a feature.

3. Does a listed rod fit?

Does it fit? checks a candidate against the agitator's profile rule by rule: the all-metal rod of the same size passes, the plastic one is rejected on its material and its diameter, the metal one with wider tabs is rejected on the span, and the listing with no measurements stays at needs evidence, its unknowns turned into questions. The verdict is one of five on the ladder, decided by the rules, never by the model. Does it fit?.

4. Find the missing piece

Find a part records what the agitator needs (all metal, a 4 mm rod, two tabs 9 mm across) with a top price. Plan the search has Muse Spark write three to six Marketplace search links you open yourself, a match pattern and the words that rule a listing out. Listings pasted back are ranked by the pattern and the price with no model call; Check fit has Muse Spark read one into fields, dated, assessed by the same rules and placed where it stands, and Ask the seller turns the open questions into a message. Find a part.

5. What should become of it?

On any item, Decide, What should I do with it?: Muse Spark writes three cards for the item and for the owner's profile, its skills rated from none to pro: keep and fix (the likely problem, the steps at that level, the parts, cost and time), reuse it (two or three builds with what else the owner has) and sell or give away, with one recommendation. Choose this keeps the choice as the item's plan and moves to the section it leads to. Decide what it becomes.

6. What could the monitors become?

Build and reuse: say what you want from a build, ask for proposals, and Muse Spark reads the inventory through tools and proposes projects from templates kept as data; only what names items you own is shown: a homework workstation on the 24 inch Dell and a smart mirror for the hall on the 18 inch one. Select the homework workstation: its items are committed, the budget line says what is left, the missing cable becomes a search, and the build pack is one document from what stands. Draw it has Muse Image draw how the finished desk could look, one image call, captioned as drawn. Record the outcome once it works. Build and reuse.

7. Sell the graphics card, give the processor away

On the graphics card, What is it worth? has Muse Spark search for listings and sales of the same model and keeps only the comparables a page backs; asking, sold, reported and estimated stay four numbers. Under Prepare to sell or give away, Listing details take the condition, what is included, the faults, the handover and the area and the asking price; Write the draft has Muse Spark write a Facebook Marketplace (US) draft from confirmed values, with the disclosures the rules append; the slot 1 processor goes as a free-pickup draft, and the older ThinkPad as one with the data-removal steps. Each field has Copy, Download photos zips them in order, and Open Marketplace opens the create page; posting is yours. Value an item, Seller flow.

8. On the phone

The Android app points at the same backend: a photo of a label read and saved into the same inventory, and a voice note held to the item, kept as your words and read back. Android app.

Running it in public

Behind an access code with allowances per visitor and per day, a spend cap and an operator's switch: Demo access.