Work you can open.

Four of these are products we built and run in production on their own domains. The other four are working demos built over fictional companies, because client names and client data are never shown.

Retail price intelligence · live in production

Price history that retailers do not publish.

Retailers reprice constantly and publish only the current number. Without a record of what an item cost last week, a sale price cannot be checked, and the retailer is the only party holding that record.

We sweep 38 retailers overnight and keep every price observed, so a cut is measured against that item's own history rather than against a manufacturer's list price. Clearance and destock checks run on their own schedules, because some prices move hourly and others move seasonally.

It has run unattended since launch. The storefront is the public front end of the price archive.

Retailers38
Overnight sweepnightly 04:00
Clearance passevery 2h, 8 a.m. to 10 p.m.
Destock detectionhourly

Open The Undersold Running unattended

Travel pricing · live in production

Discounts on rentals still empty days before check-in.

Owners often discount a vacation rental that is still empty four days before check-in. The booking platforms do not highlight those discounts to travelers.

We sweep check-ins zero to five days out across drive-from-Charlotte markets, track price over time per property, and rank on the size of the actual cut rather than on the headline nightly rate. A property with a consistently low rate does not rank as a discount.

Three full sweeps a day, with a lighter rescore every two hours to catch cuts made between them.

Booking window0–5 day check-ins
Full sweep3× daily
Hot-pass rescoreevery 2h
Focusdrive markets

Open It's Giving Gone Running unattended

Public-sector data · analysis & reporting

A long-term record of state office wait times.

A state agency published current wait times and overwrote them continuously. That served someone leaving the house in ten minutes, but it could not show whether waits were getting longer, which office was slowest, or whether a policy change had any effect.

We built a collector that samples on a schedule through the operating day and keeps every reading. The analysis layer sits on top: what the wait was at closing, what the day's peak was, how one office compares to itself last quarter.

Because the sampling rate changed partway through, the analysis aggregates by day before it aggregates across days, since pooling the raw readings would overweight the denser period. The dataset is the only longitudinal record of its kind.

Samplingevery 10 min, weekdays
Retentionfull history, no overwrite
Coveragestatewide
Outputpublic archive + reports

Open Seen NC Running unattended

Consumer transparency · design & build

How long each unit has been listed for sale.

A listing shows the price. It does not show that the same machine has been on the same lot for eleven months.

We built a public consumer site over the same inventory data that powers the private dealer audit. It runs as a separate process that reads through its own data layer and cannot reach private audit findings.

The site is open to search engines, so a buyer in a showroom can find the model in front of them by searching on a phone.

Coveragestatewide
Readspublic listing facts only
Private findingsunreachable
Searchindexable

Open Listed Since Running unattended

Inventory reconciliation · fabricated demo

A dealer advertising 180 units and holding 149.

Ninebark Motor Group is not a real company. Every unit, price and photograph in this demo was fabricated. The findings were computed by the same audit engine that runs against real dealers, pointed at the invented inventory.

Dealers routinely have three sources of truth for inventory (the website feed, the DMS, and the actual floor) and no way to compare them. Units sold weeks ago keep drawing leads. Units physically on the lot never made it online.

The audit grades every unit with explicit provenance: confirmed on the floor, catalog-only, or ambiguous. The absence of a VIN is not treated as proof that a unit is missing. Eleven of Ninebark's units are in the ambiguous group, and the audit does not classify them either way.

It also finds units one store publishes and its sister store does not: eight machines listed on the euro storefront never appeared on the sister site. These were found from outside, with no inside access.

Advertised online180
Confirmed on floor149
Phantom listings31
Ambiguous, not judged11
Cross-publish gaps8
Honesty score91.4 / 90.4

Open the Ninebark demo Simulated data

Distributor catalog & storefront · fabricated demo

A distributor catalog rebuilt nightly from supplier feeds.

Halloway Educational Technologies is not a real company. This demo runs the real catalog pipeline over a de-identified 3,814-product slice. Distributor cost, the original listed price and the MAP floor are deleted rather than adjusted (an adjusted cost is still a cost), and every price shown was moved deterministically. The figures on the demo's own pages are illustrative rather than computed.

A dealer catalog of tens of thousands of hand-keyed items drifts over time. Lines get added and never retired, a manufacturer winds down and its products stay quotable, and a price set one year gets honored the next. A catalog that size cannot be kept current by editing items one at a time.

The storefront is not edited by hand. It is rebuilt overnight from distributor feeds: normalized against a vendor-agnostic product schema, floored at MAP, and enriched so the catalog is searchable by more than a part number. Each run writes a changelog of what moved.

When a distributor feed changes format without warning, the pipeline fails loudly and keeps the last good catalog rather than publishing stale prices.

Products in this demo3,814
Departments14
Distributor cost shownnone (deleted)
Pricessimulated
Rebuildnightly, unattended
Manual updatesnone required

Open the Halloway demo Simulated data

Payments, shipping & job tracking · fabricated demo

Payment, prepaid shipping and job tracking for a restoration shop.

Ferrule & Grip is not a real company. Every customer, address, order and price in this demo is invented, and the addresses are IANA-reserved and cannot reach a person. The photographs show real restoration work and are used for illustration. The instance carries no payment, shipping or email credentials, so it cannot charge a card, buy a label or email anyone.

A one-person restoration shop was selling through a marketplace that offered no prepaid inbound label, no way to quote work that can only be priced once the item is on the bench, and no answer to “where is my set” except a phone call.

The whole path is one system. The customer picks services from an editable menu and pays, with the shop as merchant of record and our cut collected as an application fee rather than a separate invoice. A prepaid inbound label is bought and emailed automatically. Each item becomes a row on a ticket, and inspection can raise a supplemental invoice that pauses the work until it is paid or declined.

A shipping label is never bought inside a checkout. It is a retrying, database-backed task. If it fails permanently, the owner cuts a label by hand, which is his current process, so a failure never leaves a paying customer without a label.

Paymentsmerchant of record is the shop
Inbound labelprepaid, automatic
Priced at the benchsupplemental invoice
Customer accountsnone (magic link)
Live credentialsnone, cannot transact
Demo jobs9, across every stage

Open the Ferrule & Grip demo Simulated data

Traceability & claims integrity · fabricated demo

The build fails if the data can’t back the claim.

Bidwell Family Beef is not a real company. There is no such ranch, family, processor or animal. It runs the same code and templates as a real client pitch, built from a second dataset. The entire site is re-skinned by editing data, with no code change.

A direct-to-consumer food brand depends on its claims: grass-finished, no hormones, single-origin, dry-aged twenty-one days. Those claims are written by whoever builds the page, and a normal publishing stack does not check whether the operation supports them.

Here the data model carries the facts and the marketing copy has to match them. Every product traces to a lot, and every lot to an animal group, a pasture, a harvest date, a processor and a hang time. Those values are computed from the records, not typed into the page.

Then a claims linter reads the copy against the data and fails the build on anything unsupported. Adding “certified organic, 100% grass-finished” to this demo stops the build with unsupported-claim: promise.6.body -- claims 'organic' but operation_type is 'UNKNOWN' and exit code 1. The page cannot be published until the claim is removed or the data supports it.

Pages38, statically rendered
Re-skin costone dataset, zero code
Unsupported claimfails the build
Traceabilitylot → group → pasture
Hang time, weightscomputed, not declared
Backendnone (static files)

Open the Bidwell demo Simulated data

Tell us what's going on.

Describe the problem or the project in a few sentences. We'll reply with next steps.

[email protected] →