Building AutomateEdge Solo — A Free, AI-Curated Learning Platform
Building AutomateEdge Solo
A free, self-paced platform that gives anyone a personalised, AI-curated study plan to learn modern AI from scratch — then tracks their progress, grades their practice work, and hands them a verifiable certificate when they finish.
This is the honest story. Why I built it, what it actually does today, what I got wrong along the way, and where it's going next.
Why this exists
The AI learning content that already exists falls into two buckets.
Static roadmaps are one-size-fits-all and drift within a quarter. The thing at the top of "you should learn this first" is rarely the right starting point six months later. Nobody maintains them.
Paid cohort programs solve the currency problem but gate on price and schedule. If you work a full-time job and live in a timezone where the live sessions fall at 2 AM, you're not the target.
AutomateEdge sits between them. The curriculum is curated from trending signals — university course catalogues, practitioner writing, current research themes — and re-evaluated every quarter. It's free. Learners enrol, get a plan tailored to their duration, level, and career goal, link their practice work as they build, and graduate with a signed credential they can show a recruiter.
What it does for a learner
You pick a duration (3, 6, or 12 months) and a level (beginner, intermediate, advanced). You get a week-by-week plan with focus areas, deliverables, curated resources split into video courses and docs/practice, and a checklist per week.
You can link your GitHub repos to the weeks they match. An AI mentor can answer questions scoped to what you're working on right now — it sees the current week's focus areas, not the whole internet. If you've told the platform you're targeting a security engineer role, the mentor anchors its examples there instead of a generic data-science framing.
When you cross 90% completion with the capstone month at 100%, you earn a Completion certificate. Finish everything and link enough repos and it upgrades to Distinction. Ship something an automated evaluator scores 8/10 or higher and it upgrades again to Honors. Each upgrade preserves your original credential ID and issue date — it's a progression, not a reissue.
The curriculum stays current by policy, not magic
Quarterly refresh runs an AI discovery pass that proposes additions and retirements against the previous cycle. The output feeds a quality pipeline — generate → review → refine → validate → score — that grades every candidate template across fifteen dimensions: cognitive progression, theory/practice ratio, project density, assessment quality, freshness, prerequisites, real-world readiness, and others.
A template that scores well becomes publishable, not published. An admin still has to click Publish. That click stamps who reviewed it and when, and that stamp is visible to learners. It's a small thing that makes the difference between "some algorithm decided this was good" and "a human vouched for this on this date". The score is a filter, not a judge.
I shipped auto-publish at first. Within two days, a confidently-wrong template had gone live. I pulled it, added the human gate, wrote a rule in my notebook: every generation pipeline gets a human button, even if the button does nothing but exist.
The certificate actually means something (a little)
Most platform certificates are lipstick on a spreadsheet. A recruiter can't tell the difference between one you earned and one you forged in Canva.
AutomateEdge certificates are signed and verifiable. Every credential ID has a cryptographic signature baked into it. Anyone with the ID can visit a public verification page and see a green "Credential verified" badge — or a red one, if the signature doesn't match. If someone writes a fake ID on their resume, the verify page will say so.
The page also shows the learner's display name (snapshotted at the moment of issue, so editing your profile name later doesn't retroactively change an issued certificate), the course title, the module breakdown, completion stats, and the issue date. It's designed to be the thing a recruiter sees in 20 seconds and either believes or doesn't.
Am I calling this a real industry credential? No. It's a portfolio proof with tamper detection, and I'm honest about that distinction. The long-term play is to position AutomateEdge courses as preparation for recognised certs and let the existing trust do the heavy lifting.
The leaderboard is the most motivating page on the site
It started as a "maybe some people want to see a ranking" idea and turned out to be the highest-engagement page I have.
Learners are ranked by XP, which is earned from real activity: checklist tasks, GitHub repos shipped, certificates earned (weighted by tier), and active weekly streaks. There are seven tiers — Apprentice, Learner, Practitioner, Builder, Engineer, Architect, AI Guru — each with its own coloured chip and a mini progress bar showing how close you are to the next one. Achievement pills decorate the page for milestones like First Task, Triple Crown (three certs), 10-week Hot Streak.
Two rules keep it honest. Distinct repos, not repo-links — linking the same repo to five weeks doesn't earn you 5× the XP, so you can't farm. Every number is a live query — there's no placeholder data, no fake users, no editorial curation. If the page says you're Top 5%, the arithmetic agrees.
There's also a help panel on the leaderboard that spells out exactly how every signal maps to XP, what each tier requires, and what triggers each badge. No mystery meat — if you want to climb, the rules are published.
What I got wrong
The auto-publish decision, covered above. If a generation pipeline ever lands work that ships to users without a human button, I have failed the learner.
Feature breadth before traction. I built an internal cost dashboard, batch refinement, a provider fallback chain, and embedding-based topic dedup before I had more than a handful of users. The right order was the opposite: get people on the site, learn what they actually need, build for that.
Gut-based picks that cost time. I spent weeks convinced one specific AI model was the right tool for a specific task because it had the strongest reputation. The measurement eventually said a different model was better for that workload. I had the evidence well before I acted on it. Lesson: trust the benchmark you ran, not the benchmark you read.
Not writing a post-mortem every session. This is it. I'm fixing it now.
What's next
- An admin UI for revoking a credential if a genuine issue surfaces.
- Timeframe tabs on the leaderboard (all-time, this month, this week) — new learners can climb the monthly ranking fast and stay engaged through their first 30 days, which is the riskiest window.
- An AI news feed and an AI jobs board. "I finished my course, now what?" is the most common question from early learners; answering it on-platform is the obvious move.
- More courses beyond the three generalist tracks. The quality pipeline is mature enough now that the bottleneck is human review capacity, not generation.
If you're thinking of building something similar
Ship the boring version first. Every layer you don't add is one you don't have to keep alive.
Build for the user you have, not the user you wish you had. A free product with ten real learners beats a brilliant product with zero.
Measure everything that isn't free, and set a hard budget cap. It's the difference between catching a mistake at a dollar and discovering it at fifty.
And write the post-mortem at the end of every session, even if the only reader is tomorrow-you.
