AppVantix

Work

Our own product, as the worked example.

We have no client names to show yet, and we will not invent any. Here is a product we built and ship ourselves, with the evidence.

Private beta

Looking Glass for SAP

A desktop command palette for SAP support teams. A person types plain words, or a transaction code and a number, and it opens the right SAP screen with the number already filled in. It labels display, change and post differently and asks before anything that can change data. It never reads from or writes to SAP.

macOSsigned and notarized; Windows build not yet run on Windows
No networkmakes no requests of its own, measured and documented
No datacollected by us; settings stay on the person’s computer
Free betafor companies, on a test SAP system, under a short agreement

The five rules, applied

What the rules looked like on this build.

01

Specified before built

A written product plan set the boundaries first: open screens, never act; no connection to SAP; change screens ask before opening.

02

Tested on data we did not tune on

Three held-out sets were written before and kept apart from tuning. The right transaction is in the top three for 92 to 95% of requests on them, against the built-in list of 142 transactions. We also report the misses.

03

Rules decide, not the search

Ranking only proposes. A fixed gate checks the person’s role, the chosen system and the Production block, and asks for confirmation before any change screen.

04

Data stays put

The app has no accounts, telemetry, crash upload or auto-update. Tests fail if network code or a dependency is added, and a network log confirms it.

05

Limits stated

It has been tested against a stand-in for SAP, not yet a live system. The Windows build is untested. Both are written on the product page.

The personTheir computerLooking Glassno network of its ownDefault browserTheir own SAPsign-in decidesone address
Data flow of Looking Glass. Nothing leaves the computer except one address handed to the person’s own browser.
Held-out test setRequestsRight answer firstIn the top three
Set B (never tuned on)4875%92%
Set C (written before the work, never tuned on)7777%94%
Set A (later used to tune)6180%95%

Measured by the product’s own test suite against its built-in list of 142 common transactions, not against a customer’s system. Real accuracy for a company comes from importing its own list.

Start here

Want this level of evidence on your project?

A short note is enough. We reply within two business days, and the first conversation is free.