01
Repetitive Questions
Admissions and recruitment teams repeatedly answer variations of the same questions about requirements, documents, deadlines, programs, and application steps.
University Admissions Copilot is a portfolio product currently in development. The concept is based on a problem I know well from international education: prospective students repeatedly need answers about admission requirements, application documents, deadlines, programs, tuition, and institutional processes, while staff need those answers to remain accurate, current, and consistent.
Current status: product design and technical planning
This is not presented as a finished or deployed product. The problem definition, target user experience, feature set, and planned technical approach are being designed now. The next phase is to build a working prototype, test it against a controlled set of institutional documents, evaluate answer quality, and refine the product based on what the prototype demonstrates.
The Product Problem
Admissions information exists. Finding the right answer is the hard part.
Universities publish large amounts of information across program pages, admissions pages, policy documents, FAQs, tuition pages, deadlines, and supporting documentation. The challenge is not simply generating an answer. It is retrieving the right institutional information, maintaining context, and responding without inventing details that the source material does not support.
02
Fragmented Information
Prospective students may need to search multiple webpages and documents before they can piece together a complete answer.
03
Accuracy Matters
Admissions guidance has real consequences. A useful assistant must distinguish between information supported by institutional sources and information it does not know.
Envisioned Experience
A simpler way to navigate admissions information.
Ask naturally
A prospective student could ask questions in plain language rather than needing to understand the structure of the institution's website or know exactly where a policy or requirement is published.
Answer from approved sources
Responses would be grounded in a controlled collection of institutional webpages, documents, FAQs, program information, and admissions materials rather than relying only on the model's general knowledge.
Show the source
Where practical, the copilot would surface the source used for the answer, allowing the student to verify the information and continue to the official page.
Know when not to answer
If the available knowledge does not support a reliable answer, the product should say so and direct the user toward the appropriate admissions or institutional contact instead of guessing.
Planned Technical Approach
Retrieval first. Generation second.
The prototype is planned around a retrieval-augmented generation approach. Rather than asking a language model to answer admissions questions from memory, the system would first retrieve relevant content from a defined institutional knowledge base and then use that context to generate the response.
01
Knowledge Sources
Collect a controlled set of admissions pages, program information, policies, FAQs, deadlines, and supporting documents.
02
Prepare Content
Organize and segment source material so relevant passages can be retrieved efficiently when a user asks a question.
03
Retrieve Context
Use semantic retrieval to identify the pieces of institutional information most relevant to the student's question.
04
Generate Answer
Provide the retrieved context to the language model and instruct it to answer within the boundaries of the available institutional evidence.
05
Evaluate
Test responses for factual accuracy, source alignment, relevance, clarity, unsupported claims, and appropriate escalation behavior.
Planned Features
Useful first. Ambitious later.
The initial version is intentionally narrow. The objective is to prove that a grounded admissions assistant can answer common questions reliably before adding more complex workflows or automation.
Admissions Q&A
- Admission requirements
- Required documents
- Application deadlines
- Application process questions
Program Information
- Program availability
- Credential and study options
- Published tuition information
- Links to official program pages
Source-Aware Responses
- Ground answers in approved content
- Surface supporting sources where possible
- Avoid unsupported assumptions
- Escalate when the evidence is insufficient
Responsible AI
Accuracy is part of the product experience.
In an admissions context, a confident but incorrect answer is a product failure. The prototype will therefore be designed around explicit grounding, constrained scope, evaluation, source visibility, and clear escalation when reliable information is unavailable.
Grounded Responses
Source Transparency
Hallucination Management
Human Escalation
Privacy Awareness
Scope Control
Evaluation Before Expansion
Development Roadmap
Build the smallest credible version first.
Phase 01
Define
Confirm the target user, priority questions, source set, product boundaries, and the quality criteria the prototype needs to meet.
Phase 02
Prototype
Build the first retrieval and generation workflow against a small, controlled institutional knowledge base.
Phase 03
Evaluate
Test real admissions questions and document where the product succeeds, fails, retrieves the wrong context, or produces unsupported answers.
Phase 04
Improve
Refine retrieval, prompts, source structure, guardrails, and user experience before expanding the knowledge base or adding more advanced capabilities.
What I want this project to demonstrate.
University Admissions Copilot is designed as a hands-on bridge between my experience in international education and the AI skills I am actively building. The goal is not to present myself as an AI engineer. It is to demonstrate that I can identify a real business problem, define a focused AI use case, understand the role of grounding and retrieval, think through responsible deployment, translate requirements into a product workflow, and learn the technical implementation deeply enough to work effectively across business, product, and technical teams.