How Alpha’s LLM layer works: rules first, model second, counselor last

Alpha uses a large language model for the parts of counseling that are about explaining and planning: turning a rule check into a study plan, and turning a statement of purpose into feedback. We built that layer on Claude, Anthropic’s model, through Anthropic’s official SDK. It is in beta. Until it is switched on for everyone, a counselor writes the plan from exactly the same inputs.

What goes in

• The student’s profile without identity: level, subject, results, passing year, English score, budget and notes. No name, phone number or email leaves the site.
• The rule check from AI Match, with every point it found.
• The destination facts we publish, and nothing else. Where a fact is not published, the model is told to say that a counselor will confirm it.

What comes out

The model has to answer in a fixed JSON structure that the API enforces: a headline, a short summary, a verdict for every destination (likely, possible or not yet) with the reason and the main risk, the gaps to close, a timeline, the documents to prepare and the questions to ask a counselor.

Because the shape is fixed, the software can use the answer instead of pasting a wall of text. Verdicts become cards on the student’s result page. Gaps become next steps, and documents become a checklist. The counselor sees the same plan next to the lead.

Guardrails

• Never promise admission, a scholarship or a visa.
• A destination the rule check marks “not yet” is never offered as a current option.
• A daily call limit. Past it, requests fall back to the counselor queue.
• Plans in English or Bangla, whichever the student chooses.
• A counselor reviews every plan before the student acts on it.