What we shipped, and what we learned shipping it.

Sightline

Problem

A shooter photographs a paper target, then does the group math and the scope correction by hand, and keeps the range log in a notebook or an app that wants an account and an ad network.

What we built

Photograph the target, tap the holes, get exact scope adjustment clicks. An armory, ammo inventory with burn-rate projection, G1 ballistic DOPE cards, and a range log with CSV and PDF export. Everything on the device; it works miles from signal.

What shipped

Version 1.0 on Google Play at $7.99, one-time. No account, no ads, no subscription. The Play data-safety label reads "no data collected" because that is what the app does.

What we learned

Before the public build we audited every permission the app requested and removed the ones it did not need. A privacy claim is only worth making if the manifest agrees with it.

LaunchPad

Problem

People preparing to enter the Air Force were piecing together fitness standards, ASVAB practice, and the pre-ship checklist from forums, PDFs, and apps that were paywalled, out of date, or both.

What we built

One app: fitness tracking against the standards, ASVAB practice with a skill map by subtest, and the checklist that runs all the way to ship day. Built by a veteran for the next people putting on the uniform.

What shipped

Free on Google Play, with a single banner ad and no subscription. The app's own site runs at launchpad.djeango.com.

What we learned

Getting the facts right matters more than the feature list. Standards change; the app has to be maintained like a reference, not a launch.

The Self-Hosted LLM Kit

Problem

Running a capable language model on hardware you own is cheaper and more private than renting one, and almost every guide about it assumes you already know how.

What we built

What started as a single homelab guide became a four-tier kit: Builder (pick parts, first run), Operator (a hardened server, three ready-made builds), Architect (Mac, PC, and server in sync), and the Full Stack (dashboard, memory, one assistant). A free Fit Check tells you what your machine can run before you spend anything.

What shipped

Five tiers on Gumroad, from free to $149, with config packs and scripts that were run before they were written up. Existing buyers got the refreshed Builder tier at no charge.

What we learned

The kit is credible because the studio runs the same stack. Every command in it carries three plain lines: what it does, what you should see, and what to do if you do not.

The AI visibility audit method

Problem

A local business owner cannot tell whether ChatGPT, Gemini, Perplexity, Google AI Overviews, or Copilot mention them, name a competitor instead, or get the facts wrong. Agencies answer with a sales deck.

What we built

A repeatable evidence method. Thirty client-approved buyer questions, run across five engines, each twice, logged out. Three measured rates (visibility, citation, accuracy), a competitor table, a cited-source tally, a technical readiness check, and a fix list split into what you can do yourself and what needs ongoing work. Every answer is kept verbatim in an appendix.

What shipped

The method, the branded report template, and the audit as a fixed-scope service. The first audits run entirely in house.

What we learned

We have no client results to show yet, and we will not publish typical-improvement claims until we have measured enough audits to mean it. The evidence is the product.

SOPSignal

Problem

Teams lose track of who owns a procedure, when it was last reviewed, and whether anyone would notice if it went stale. The fix usually means moving the procedures themselves into yet another tool.

What we are building

A metadata-only layer: source links, owners, review cadence, and decision history. Never the procedure contents, never credentials, never customer data. Point it at where your procedures already live.

Where it stands

Local prototype with fictional data. 104 unit and 122 integration tests pass. Builds run on a schedule in GitHub Actions, one pull request per run, and the build never merges its own work; a person does.

What we learned

A ledger that cannot be edited by the thing it judges is worth more than a faster build. The workflow that runs the build is not allowed to touch the workflow file.