Skip to content
AVAILABLE NOWMUMBAI · UTC+5:30
Aniket VishwakarmaStill online

NOTE 03 · 30 Aug 2026

A prediction students can argue with

5 min readML · Product · Explainability · EducationFrom the Student Placement Prediction record

A placement probability is only useful if a student can do something with it. A single number that says 62% is not feedback; it is a verdict. The product needed to show the signals behind the number and the next action that would move it.

Student Placement Prediction serves students, mentors, training and placement officers, and admins. Each role sees a different slice of the same model, which made the output contract the most important design decision.

Two modes, one contract

Students can enter academic and interview details or upload a resume. Those are different inputs, but they should not produce two different products. Both modes return the same prediction shape: a probability, a band, ranked drivers and a model version.

That contract let the interface stay stable while the model changed underneath. It also made the resume path testable against the questionnaire path instead of becoming a separate feature with its own drift.

tsfield note 03
type Prediction = {
  probability: number;      // 0..1
  band: "low" | "medium" | "high";
  drivers: Array<{
    label: string;
    weight: number;
    direction: "up" | "down";
  }>;
  source: "questionnaire" | "resume";
  modelVersion: string;
};

Make the rubric visible

The ATS score is deliberately not a black-box classifier. It is a 0–100 rubric built from keyword density and formatting signals, and the interface shows which signals moved the score. A student can disagree with a specific input instead of arguing with an oracle.

That choice trades sophistication for legibility. For this audience, it is the right trade: a transparent heuristic that a student can act on beats an opaque model that is marginally more accurate.

Auth at the session edge

Four roles, one session model. Firebase Auth handles identity, but authorization lives on the server with HTTP-only session cookies, CSRF protection and role checks at the route boundary. Client-side role checks are a UI convenience, not a security control.

The model is a product

The current model is trained on sparse, noisy academic data. The next step is not a larger model; it is versioned training runs, drift monitoring and an evaluation set that reflects the students actually using the product. A prediction is only as honest as the data behind it.

  • Log model version with every prediction.
  • Track calibration, not only accuracy.
  • Measure whether students act on the feedback.
  • Review the rubric with the people who use it.
If a student cannot see why the number moved, the number is not feedback.