Students
Fast, reliable answers with visible sources, at any hour.
human-supervised ai
An AI helpdesk that only answers when it can prove it
A student assistant grounded in the official documents of a degree programme — five checks, three possible outcomes, human approval and automatic re-checking when documents change.
Context
Students of a degree programme ask the same questions every day — enrolment, study plans, internships, labs, exams, credit recognition, exam dates. The answers exist, but they live inside regulations and documents dozens of pages long.
On these topics a wrong answer is not a small inconvenience — a wrong date or an invented requirement can cost a semester. A generic chatbot, capable of "hallucinating" plausible content, was out of the question. So the design question became a different one — how do you build an AI assistant an institution can trust?
The problem
- Repetitive questions absorbing the time of the people who support students.
- Answers scattered across long documents, hard to consult in a hurry.
- Generic language models make things up — unacceptable for dates, amounts and requirements.
- Documents change during the year — and yesterday's correct answer can become wrong today.
- Student questions may contain personal data that must never reach an AI system.
Users and goals
Service supervisors
Approving, correcting or rejecting AI proposals from a single queue, and keeping an eye on how the service is going.
Programme offices
Receiving only personal cases and uncovered topics, not questions answered by regulations.
- Zero invented answers — every statement must rest on a cited official document.
- No student personal data processed by the model.
- A person always has the last word, through a sustainable approval flow.
- Answers that stay correct over time, even when documents change.
The process
- Analysis
Mapping recurring questions and the official documents that cover them, with the people who support students daily.
- Pipeline design
Definition of the five checks and three outcomes — the heart of the system is the question's journey, not the model.
- UX/UI
Two interfaces — the student chat and the control panel with approval queue, questions, answers, documents and trends.
- Release and tuning
Real operation under supervision, refinement of recurring cases and built-in visual documentation.
Architecture and modules
The system consists of the student chat, the supervision panel and an engine that treats every question as a case to be processed — with checks in a fixed order and a tracked outcome.
Student chat
An essential interface with suggested questions, openable citations and clear instructions about what not to enter.
To approve
The queue of pending proposals — approve, correct or reject, with answer and document side by side.
Questions and Answers
A browsable archive of what was asked and the library of approved answers.
Documents
The official sources monitored by the system, with confirmation of new versions.
Trends
Questions received, outcomes, most requested topics and student feedback.
Activity and Users
Operation tracking and access management.
The key flow
Five checks, three outcomes
Every question goes through five checks in this order. No step can be skipped, and the outcome is always one of three — tracked.
- 1 · Personal data
Emails, student IDs and names are removed before any processing.
- 2 · Personal case
If the question concerns an individual case, it never goes through the AI — it goes to the right office.
- 3 · Coverage
If the topic is not in the documents, the system says it does not know. It never improvises.
- 4 · Official documents
The answer is written by reading the regulations, citing every source.
- 5 · Number verification
Dates, amounts and credits must exist in the documents — otherwise the answer is discarded.
Three possible outcomes: Resolved (a human-approved answer already existed — it arrives instantly, without AI), From the documents (a new proposal with citations, entering the approval queue) or Referred (the system does not answer and directs the student to the right office).
When a document changes
The most delicate point of a document-grounded system is time. Every day the system re-checks its sources.
- Daily re-check
Sources are re-downloaded and compared with known versions.
- Suspension
If a document changed, the answers depending on it are automatically suspended.
- Revalidation
Answers that pass the re-check go live again; the others return to the human approval queue.
Main features
Openable citations
Every answer shows the documents it comes from, one click away.
Approve · Correct · Reject
Supervision is a three-button flow — correcting a proposal fixes and approves it in one move.
Instant answers from the already-approved
Recurring questions receive the existing approved answer, immediately and without AI.
Service trends
Volumes, outcomes, top topics and feedback — the service is measured, not assumed.
Documented recurring cases
The recurring supervision situations are explained right inside the panel.
Voice input and suggestions
Suggested questions to get started and voice input for those who prefer not to type.
Student assistant Answers from official documents
How many internship hours are required in total?
The programme defines a total number of hours in its regulations. You can find the per-year breakdown in the table of the document cited below.
📄 Internship regulations · art. 4 ✓ Answer approved by a personDemonstrative reconstruction of the interface, with fictional data.
UX decisions
The chat declares its limits
Always visible under the input — do not enter personal data, answers do not replace official documents. Trust starts from a clear perimeter.
The panel explains the system, not just the data
The first section of the supervision panel is a visual guide to how the service works — supervisors must understand the mechanism to trust the outcomes.
Transparent outcomes
Every question has an explicit, browsable outcome — even "no answer" is a designed outcome, with the right referral.
Correcting is faster than redoing
The correction flow turns supervision from a bottleneck into a quick pass.
Security and privacy
- Emails, student IDs and names removed before any AI processing.
- Personal cases are never processed by the model — they go to the competent office.
- Questions are logged anonymously to improve the service.
- Every published answer has passed human approval.
- Panel access with dedicated accounts; the service runs on controlled infrastructure.
Challenges and solutions
Challenge. Language models tend to produce plausible but unfounded answers.
Solution. A double constraint — retrieval over official documents only, plus separate verification of numbers. Without grounding, the system prefers not to answer.
Challenge. Documents change during the academic year, silently.
Solution. Daily re-checking with automatic suspension and revalidation of dependent answers.
Challenge. Students write in a hurry, with typos and wildly different phrasings.
Solution. Questions are normalised and matched to existing approved answers when the topic is the same.
Outcomes
- In production for the programme's students, with active human supervision.
- Every published answer is grounded in cited documents and approved by a person.
- Personal cases and uncovered topics are routed to the right office instead of receiving improvised answers.
For confidentiality I do not publish service metrics; the image shows the real panel with the university's branding removed.
Possible evolutions
- Extension to other degree programmes with separate document bases.
- Proactive reminders on recurring academic deadlines.
- Additional channels (a widget in the programme website) on the same pipeline.
Does this process sound familiar?
Tell me about it: the first orientation call is free of obligations.