In November 2025, Mereka’s learning platform lived inside Kajabi, a hosted course tool. Eleven months later, our courses, learner accounts, enrolments and certificates sit on our own platform, built on the open-source Open edX.
We did not hire a development agency to get there. AI coding agents such as Claude Code and Codex co-wrote nearly half the code.
Key takeaways
- Small, well-defined tasks with a clear test of done are what make AI coding agents productive.
- Write every mistake down as a rule, so agents learn once and do not repeat it.
- Keep people on the decisions that matter: what to keep, what to delete and what touches live data.
- Count before you claim. Reconciliation between old and new systems caught the errors that mattered.
The problem
Kajabi served us well when the academy was one brand running a handful of courses. By 2025 it was carrying programmes for several organisations, including Mereka and Biji-biji Initiative, and well over 100,000 learner accounts. Three things pushed us to move.
- Cost without control. We paid USD 649 a month and still could not change how the platform worked.
- Integrations we could not build. When a partner wanted a full API connection to their learner platform, the team concluded it was not feasible “due to the backend limitations (Kajabi).”
- One academy for many clients. Each partner organisation wanted its own branding, catalogue and sign-in. A single hosted academy could not separate them.
It’s costing us quite a bit monthly, moving it earlier will give us some savings.”Ambika, on the migration timeline
Open edX, the software behind edX and many university platforms, usually needs a dedicated engineering team to run, customise and migrate years of data into.
The solution
Gurpreet, who leads our platform engineering, made the first commit on 6 November 2025. By February 2026 the team was shipping more than 1,700 changes a month with the same headcount. What changed was how the work was organised for AI agents.
Work is broken into small, checkable tasks
Every piece of work is written as a short task with a clear test of done, in a task tracker the agents can read and update.
“let agents burn down well-defined work in larger verified tranches.”Gurpreet, platform engineering
Agents write the code; tests check it
Claude Code and Codex draft features, migration scripts, fixes and tests. The platform now has more than 1,200 automated test files, most of them written alongside the features they protect. Several agents can work at once, coordinating through a shared message board so they do not edit the same files.
Hard-won lessons are written down once
When something goes wrong, the fix goes into a written procedure (we call them skills) that every future agent reads before starting similar work. Our Kajabi migration skill, for example, lists the traps that cost us time the first time round, so no agent repeats them.
People decide; production is gated
Agents propose and people decide which courses to keep, delete or hide, and every decision is recorded with a number and a date. Changes reach the live platform in three stages: a development copy, a test copy, then the live platform. Production data changes follow one rule: the agent plans it, and a person approves it before it runs.
AI checks the platform too
QA agents walk through the live platform every day as a learner would, and file what they find as issues for the team.
The migration itself ran in waves. Agent-written scripts moved courses, accounts, enrolments and completions across, then compared counts between Kajabi and the new platform.
“Hi team, just want to announce that we have successfully migrated all courses from kajabi to our new LMS!”Miranda, 15 September 2026
The results
| Metric | Before (Nov 2025) | After (Oct 2026) |
|---|---|---|
| Platform | Hosted Kajabi, USD 649 a month | Our own Open edX platform |
| Courses | On Kajabi | 136 on our platform, 66 migrated from Kajabi |
| Learner accounts migrated | None | 166,775 accounts, 587,407 enrolments |
| Certificates carried over | None | 13,949 |
| Code changes | 0 | 17,000, 46% co-written by AI agents |
| Organisations with their own branded academy | 1 shared academy | 3 separate branded academies |
Source: Mereka’s tech team, October 2026
Part of the saving has already arrived. Cleaning up inactive Kajabi contacts brought the bill down from USD 649 to USD 499 a month, about USD 1,800 a year, while we finish the move. Kajabi takes its last new sign-ups on 31 October 2026, and the subscription ends once paid courses have a payment route on the new platform. We have not yet put a net figure on it, because the servers and AI tools behind the LMS are shared with other Mereka platforms.1
The bigger change is ownership. New programmes now launch on a platform we control. The ESG Sustainability Leaders Programme (ESLP) from Biji-biji Initiative was the first programme onboarded. When a partner asks for a feature, the answer is no longer “the platform doesn’t do that.”
Lessons and what’s next
Agents made mistakes, and our checks were built to catch them. Reconciliation counts between Kajabi and the new platform flagged every error, from empty placeholder lessons in early course builds to certificates attached to the wrong course. Without those checks, the speed would have moved the errors across too.
Next is bringing learners back. Most of the migrated accounts were dormant on Kajabi, so onboarding and re-engagement are the next piece of work. So is the payment gateway that lets us switch Kajabi off for good.1
Sources and methods
1 Mereka’s tech team, October 2026
Figures come from Mereka’s tech team: the migration reconciliation counts between Kajabi and the new platform, the platform’s code history and the Kajabi invoices. Measured in October 2026.
