Case study · AbroBot
How MNB Research built AbroBot’s AI.
AbroBot is a study-abroad platform for Indian students. We built its AI the other way round from most: the data first, then a counsellor that is only allowed to answer from it. Today it answers from official records for 3,949 institutions, and every figure arrives with the year it belongs to.
The problem
Study-abroad advice came from memory, or from commission.
A family can commit four years and upwards of forty lakh rupees on the strength of a few numbers: what the course costs, how hard it is to get in, what graduates earn. Those numbers usually arrive without a source or a year.
Advice from memory
Counsellors quote fees and admit rates from experience. General chatbots quote them from training data of unknown age. Neither tells you which year the figure is from, or where to check it.
Advice paid for by universities
Many agents earn a commission from the universities they place students into, and the student is rarely told which ones pay. AbroBot takes no university commissions, so its advice had to stand on evidence instead.
The right number, read wrongly
A US acceptance rate is a first-year undergraduate figure. Quoted to a master’s applicant without that label, an accurate number becomes misleading advice.
What we built
A database first. Then a counsellor bound to it.
Five pieces of work, each of which a student can feel in a single conversation.
The AbroBot Database
Version 2.2.0, built 13 November 2026. 28,759 institutions in 246 countries, 87,295 US degree programmes and over 2.1 million data points.
Official sources only
Government data, national surveys and open registries in five countries, plus World Bank indicators and Wikidata linked through ROR IDs.
Grounding rules
Quote the official figure with its year. Say so when a university or figure is missing. Label US admission rates as first-year undergraduate.
Conversation memory
The counsellor remembers the last 8 turns, and universities named in recent messages are looked up in the database again, so follow-ups stay accurate.
Locked down
The database cannot be read publicly. Every read goes through the app’s server.
IPEDS, US Department of Education: admissions, tuition, enrolment, graduation and completions by programme
College Scorecard, by institution and by field of study
National Student Survey 2026; LEO graduate earnings from tax records, tax year 2022-23
QILT Student Experience Survey 2025 and Graduate Outcomes Survey 2025, with separate figures for international students
HRK Hochschulkompass, the official register of institutions
NIRF 2025
World Bank indicators; Wikidata (CC0) linked via ROR IDs
Government visa living-cost requirements for the UK, Canada and Germany
Proof
One answer. Every figure checkable.
A student asks about Northeastern University for a master’s in computer science, then follows up on tuition. These are the figures the counsellor works from, each with its source and year, and each one public.
What the counsellor says
- Northeastern awarded 833 master’s degrees in computer science in 2022-23, and 608 went to international students
- Graduate tuition was $27,728/yr (2023-24)
- Its 5.7% acceptance rate (Fall 2023) is the first-year undergraduate figure, not a master’s admit rate
- Asked “what about its tuition?”, it knows “its” means Northeastern
What it will not do
- Round a figure into “around $28k” and drop the year
- Call an official figure modelled, or an estimate official
- Present 5.7% as the student’s chance of a master’s place
- Guess a figure the database does not hold. It says it is not there
Figures from AbroBot Database v2.2.0. The wording of a live answer varies; the figures, years and sources do not.
Results, today
What is live, and what is held for later.
We separate the two because a case study that blurs them is a brochure. As of 14 November 2026:
Live in the app
- Official records for 3,949 institutions in the counsellor chat and university search
- Each answer quotes the figure with its year
- Search result cards for 3,279 institutions show official tuition, acceptance and a highlight such as median earnings
- Follow-up questions across the last 8 turns
- A database that cannot be read publicly
Held in the AbroBot Database
- 28,759 institutions in 246 countries
- 87,295 US degree programmes, with degrees awarded in 2022-23
- How many of each programme’s degrees went to international students
- Programme-level earnings for 41,784 programmes
- Over 2.1 million data points in all
What’s next
Programme-level answers, wired into the chat.
The programme data is already in the database. The next step is connecting it to the counsellor, so that a student asking about a specific course gets that course’s numbers, not only the institution’s. Until that ships, the chat answers from institution-level official records for 3,949 institutions, and says so.
Degrees, by programme
How many degrees each of the 87,295 US programmes awarded in 2022-23, and how many of them went to international students.
Earnings, by programme
Programme-level earnings for the 41,784 programmes that have them, so a student can compare the course, not only the university.
The same rules
Every new figure arrives under the rules that hold today: official source, year attached, and a plain answer when the record is not there.
For your organisation
The method carries over.
AbroBot is a study-abroad product, but the work behind it is the work most enterprise AI needs: assemble the records, bind the model to them, and make it say when it does not know.
Data before model
The counsellor quotes from a versioned database, not from what a model remembers. Your documents and records deserve the same treatment.
Rules it cannot talk past
Grounding rules written once and applied to every answer, not left to the phrasing of each question.
An honest gap
When the record is missing, the system says so. For a regulated business, that is the difference between a tool and a liability.
Build an AI your customers can check.
A 30-minute call with the founder, then a written proposal that says what the system will answer from, how it will be tested, and what it will refuse to guess.