Selected work

Higher education · AI · 2025

An admissions copilot that handles the first hour.

Voice and chat in one channel, three languages, trained only on the institution's own admission policies, and built to hand off to a human counsellor the moment the stakes rise.

ClientConfidential. India.
SectorHigher education, admissions
ServicesAI architecture, conversation design, voice systems, evaluation, counsellor training
Timeline16 weeks to live
EngagementOngoing. Second admissions cycle.
The admissions copilot handling an applicant question.

A stylised recreation of the product. Names, figures and documents are illustrative.

I  /  Context

The same fifteen questions, answered all day.

Counsellors spent most of their time on the same fifteen questions. Eligibility, fees, hostel availability, curriculum, visa paperwork. Important questions, answered the same way every time, in three languages, eleven hours a day. The people who should have been having the harder conversations were doing first-hour intake at scale.

Leadership wanted the easy questions handled by a system, with a clean handoff to a human the moment judgment was needed. They did not want a chatbot. They wanted a copilot that could speak, listen, take notes, hand off, and never invent an answer.

II  /  Approach

Sixteen weeks. Built on refusal, not confidence.

The principle from day one: the agent's first job is to know what it does not know. Every conversation, every evaluation and every training session followed from it. The hardest engineering was in the refusals.

01

Policy ingestion

We spent a week with the head of admissions and catalogued every brochure, FAQ, policy memo and email template from the previous two cycles. The agent can answer only from that corpus. Anything outside it is an automatic handoff. General world knowledge was left out on purpose.

Weeks 1 to 4
02

Conversation design

Conversations in English, Hindi and one regional language. Voice and chat in the same flow, switchable mid-conversation. Every exchange logged, every handoff annotated, every refusal explained. Counsellors were trained on what the agent would and would not do, and signed off on each flow before launch.

Weeks 4 to 8
03

Build and evaluate

We built the agent, the voice layer, the handoff and the audit log, then evaluated against three hundred real questions from previous cycles. We refused to ship until the refusal-on-uncertainty rate passed ninety-two percent. Better that the agent fetches a counsellor than guesses.

Weeks 8 to 14
04

Live, then iterate

Released two weeks before the admissions cycle opened. The first cycle handled eighteen thousand applicants in thirty days, and senior counsellors spent their time where they add value. The second cycle is in flight, with refinements drawn from real handoffs.

Weeks 14 to 16
III  /  Outcomes

Counsellors back to doing counselling.

Applicants get faster answers, in their own language, at any hour. Every conversation is auditable, every handoff is clean, every refusal is logged. The institution now has a complete record of how its admissions process behaves at the front line, which it never had before.

18,000 Applicants handled in the first thirty days
3langs English, Hindi and one regional language
92% Refusal-on-uncertainty rate before launch
Audit complete Every conversation logged, every handoff annotated, every refusal explained
The counsellors stopped answering the same fifteen questions. Now they answer the ones that matter.
Head of admissions · Higher education institution · India