04 — PERSONAL PROJECT

Smarter access
to knowledge.

UniChat: a full-stack AI study adviser for Swedish higher education. Describe a goal, subject or city and get ranked programme recommendations with grounded explanations.

VIEW THE CODE

Type: AI study adviser
Status: Personal project
Period: 2026
Tools & technology: Next.js, TypeScript, Tailwind CSS, Python, FastAPI, PostgreSQL, Redis, Qdrant, OpenAI embeddings

PROJECT / 004AI STUDY ADVISER

UniChat start page with filters for level, language and study pace next to the heading Få programförslag som känns relevanta

UniChat on a phone, with the introduction and filter button

UNI-CHAT-MU.VERCEL.APP

01Context

Choosing a programme is a big decision built on scattered information. General chat assistants are fluent but vague about specific courses. UniChat explores how far you get when an adviser is grounded in a defined set of sources.

02My role

Personal project: idea, architecture, front end and back end.

03Process

Crawl and normalise programme data, store it in PostgreSQL and index embeddings in Qdrant.
Enrich each message with session context, extracted constraints and active filters such as city, level, language and pace.
Retrieve with vector search, then apply deterministic guardrails, deduplication and reranking for hard constraints.
Generate explanations only from the retrieved programme metadata.

—A closer look

UniChat answering “I want to study something with IT in Stockholm” with ranked matches and reasons for each — Ranked matches, each with the reasons it was picked, taken from the source data.

Conversation view with suggested starting questions and a city filter — Start broad, narrow down fast: suggested questions and live filters.

UniChat header and filter panel — Filters act as hard constraints, applied in code before ranking.

04Outcome

A working full-stack application with open source code: closer to a constrained recommendation engine with RAG than a generic chatbot.

05What I learned

Hard constraints belong in code, not in the prompt.
Retrieval quality matters more than the model.

06Tools & technology

Next.jsTypeScriptTailwind CSSPythonFastAPIPostgreSQLRedisQdrantOpenAI embeddings