From a spreadsheet workflow to an AI assessment platform with about 109,000 completed attempts from roughly 48,000 candidates. I lead product and am its primary engineer.
- Contribution
- Lead AI Product Manager, Operations
- Organisation
- Skilladder AI
- Period
- May 2025 - Present
- Completed attempts
- about 109,000
- Candidates
- about 48,000
- Backend services monitored
- 128
- Faster deployment
- 85%
The problem
Skilladder runs skills assessments for hiring across Africa. When I joined, assessments ran on spreadsheets and manual coordination. Every engagement needed hands-on setup, results took real effort to compile, and nothing about the process could absorb a national programme.
What I built
A multi-tenant assessment platform, with me as its primary engineer. It covers a creator studio, AI assessment generation, grading of open-ended answers with retrieval over proprietary data, and proctoring with anti-fraud telemetry. For Nigeria's 3MTT programme I built the full delivery pipeline: the test-taking flow, proctoring, three synchronized grading engines and job matching.
Around that core sit a job board that imports jobs from PDFs, with a person reviewing each import, employer talent matching, and a client API with more than 60 endpoints and webhooks, shaped with two customer design partners.
When candidates were dropping off, I investigated and shipped an accessibility rebuild in three days, improving screen-reader support, keyboard navigation and use on low-end Android phones. Read-aloud questions with Nigerian accent voices and extended time enforced on the server sit alongside it.
Underneath, monitoring wraps all 128 backend services with automatic error classification and an admin triage dashboard, and the platform moved behind Cloudflare without disrupting email deliverability.
The stack
- React
- Vite
- Supabase
- Railway
- Cloudflare
- Claude Haiku
The hard bug
Email lost during 3MTT
During Nigeria's 3MTT programme, results and notifications went out one email at a time, and some never arrived. At national scale, a send that fails once and is forgotten is a candidate who never hears back.
I rebuilt sending as a durable queue. Each email is written to the database first and sent by workers from there, so a failed send is retried instead of lost, and delivery no longer depends on a single attempt.
The result
The platform became the company's primary product. It has served 32+ enterprise clients, including Cowrywise, ALX, Honeywell Group and Kuda, and six of them are recurring clients. Across the platform there have been about 109,000 completed attempts from roughly 48,000 candidates.
Deployment time fell by 85% against the prototype process. AI grading, rolled out through two weeks of shadow testing with confidence scores and one-click human override, cut grading turnaround from days to hours, and moving it from Claude Sonnet to Claude Haiku cut its cost from about $30 to $8 a week.
After the accessibility rebuild, completion improved and a client recovered candidates it had lost.
From a platform to a business.
32+ enterprise clients, including six recurring clients.
Business outcomes confirmed by Ayodeji, October 2026.See it
How a candidate's answer becomes a result an employer can use. The AI grades, a reviewer can always overrule it, and delivery is built not to drop results.
Take the assessment
A test-taking flow with proctoring and anti-fraud telemetry, read-aloud questions and extended time enforced on the server.
Bring in context
The question, the rubric and source material retrieved from proprietary data, so the grader judges against defined criteria.
Grade with AI
Rubric-linked scores with a rationale that can be checked against the answer. The 3MTT programme ran three synchronized grading engines.
Keep a person in the loop
Every result carries a confidence score, and a reviewer can override any grade with one click.
Deliver
A client API with more than 60 endpoints and webhooks, reports with byte-exact legacy parity, and email sent from a durable queue.
Underneath. Monitoring across 128 backend services with automatic error classification and an admin triage dashboard, with the platform behind Cloudflare.
Decisions that shaped it
Earn trust before automating
AI grading ran in shadow for two weeks before it graded on its own. Confidence scores and one-click override keep a person in control, and grading turnaround fell from days to hours.
Pay for the model the task needs
Grading moved from Claude Sonnet to Claude Haiku, cutting its cost from about $30 to $8 a week. Confidence scores and one-click override stayed in place, so a person can still catch a grade that looks wrong.
Treat fairness as a feature
Read-aloud assessments with Nigerian accent voices, extended time enforced on the server rather than in the browser, and a mobile fix that restored fair retakes for 1,300+ affected candidates.
Design the API with its users
I shaped the assessment API around triggering assessments, retrieving results and webhooks, built the first endpoints, and iterated on them with two customer design partners.
How Skilladder AI turns assessment answers into evidence.
Illustrative walkthrough · synthetic example
Start with the evidence.
An answer only makes sense against the question, the rubric and the relevant source material. Retrieval brings that context into the assessment workflow.
Give the grader the criteria it needs before asking it for a judgment.
A candidate explains how they would investigate a drop in customer retention.
Question + submitted answer + rubric + retrieved context
A model of the workflow, not a live grading run or a candidate record.
Keep exploring
Orderflow