Case 03 · 2025–2026 · Deployed
Student Placement Prediction
Placement readiness, ATS scoring and an AI counsellor that give students feedback before the drive, not after it.
01 / The brief
Placement readiness, ATS scoring and an AI counsellor that give students feedback before the drive, not after it.
Why it existed
Students usually find out they are not placement-ready after the opportunity has passed. A prediction without an explanation cannot change the outcome.
What shipped
187+ active students and 6.7k tracked events. A transparent 0–100 ATS rubric shows exactly which signals moved the score, across four role dashboards.
02 / Context
The constraint came first.
Students usually find out they are not placement-ready after the drive. This platform gives them a probability estimate, an ATS score and a clear next action while there is still time to act.
- Academic and resume data is sparse and noisy.
- A prediction is useless if the student cannot see why it was made.
- Four roles needed separate access without fragmenting the auth model.
03 / What I owned
My hands were on these.
- End-to-end Next.js application and the Scikit-learn prediction pipeline
- A shared prediction contract for questionnaire and resume-based modes
- The transparent ATS rubric and its visible signal breakdown
- Firebase Auth with HTTP-only session cookies and four-role RBAC
- Gemini integration with a Cloudflare Workers AI fallback
04 / Decisions
Four choices that mattered.
- Two prediction modes, one contractQuestionnaire and resume-based predictions share the same output shape, so the interface never forks and the model can be swapped independently.
- A transparent ATS rubricThe 0–100 resume score is keyword density and formatting, not a black-box classifier. Students can see exactly which signals moved the score.
- Auth at the edge of the sessionFirebase Auth with HTTP-only session cookies and CSRF protection, rather than client-held tokens and client-side role checks.
- AI with a degraded pathGemini powers the career counsellor; Cloudflare Workers AI is the fallback when the primary provider is unavailable.
05 / Architecture
How the pieces fit.
06 / Evidence
What can be measured.
Metrics state their source on the receipts ledger.
07 / Reflection
What broke, and what I would change.
I would version the training runs and monitor drift instead of treating the model artifact as static. The prediction is only as honest as the evaluation set behind it.
08 / Stack and fit