For most of my career, "build it ourselves" was the expensive answer. A small company bought an ERP, licensed a course-authoring tool, and paid for a simulation package, because building any of them meant a team and a year. Working alone with AI coding agents on Virabelo, I found that calculation had changed. These are the first four tools I built rather than bought, in the order they came.
1. Our own ERP
Instead of buying an ERP and bending the business around it, I started from ERPNext, an open-source ERP, and built a custom app on top of it. It became the system of record behind the drone platform: geofences, training scenarios and their execution history live there, and the operations software reads and writes them through the ERP's own methods.
That custom app grew to 266 commits between October 2025 and October 2026. The point isn't the commit count. It's that the ERP fits the business, rather than the other way around, and changing it is a code change we control, not a vendor ticket.
2. A SCORM course generator, with AI voiceover
Drone operators need training records that hold up, and the standard way to deliver tracked training is SCORM. The commercial tools that generate SCORM courses quoted more than $10,000. So I built our own pipeline. Generative AI drafts the course content and voices the narration. A converter I wrote packages the slides, audio and quiz questions as a SCORM course that runs inside our learning management system and reports scores back.
A new training class now costs the time it takes to review it, not a license. The review is the part that still needs a person, and it should.
The converter is now free for anyone to use: download it here.
3. Virabelo operations tools
The core of Virabelo is the software an operator uses to run drone missions: map-based mission planning with compliance checks as you plan, flight control over the MAVLink protocol, live telemetry, asset tracking for aircraft, batteries and docks, crew training status, and an audit trail of every change. It is multi-tenant, so each company, site and mission sees only what it should.
The main repository has more than 1,000 commits since October 2025, most of them written by AI agents under my direction.
4. An airspace simulator that grades autopilots
Flying many drones beyond visual line of sight means proving how an autopilot behaves when another aircraft gets close. Testing that with real aircraft is slow and risky. So I built a simulator that runs many drones at once in shared airspace, plays out scripted encounters, and scores how the autopilot responds.
The simulator and the autopilot code it tests have grown to about 1,700 commits since April 2026. It runs a scenario interactively or many in a batch, which turns "I think it handles that" into a scored result.
What changed, and what didn't
What changed is the cost of building. All four were built by one person directing AI agents, alongside everything else a small company has to do. What didn't change is the cost of being wrong. Each tool still needed requirements, tests, review and a way to know it worked. That is where most of the effort went, and it is the part most teams skip when they first adopt agents.
If you run engineering programs, the question isn't whether your team can build more with AI. It's whether your program can tell what got built, whether it works, and what it cost.
This is the work Loomforth does with program teams. If you're weighing build versus buy on a tool of your own, email me.