UNIVRSLL Office Assistant
Upload your business documents, then ask questions and get answers with citations back to the exact source.
The parts that matter.
Upload and organise
Drag-and-drop for PDF, DOCX, TXT, XLSX and images, with type and size validation, live processing status, and full metadata.
Semantic search
Search by meaning rather than keyword. Documents are chunked and embedded into pgvector, queried by cosine similarity over an HNSW index.
Grounded answers
Ask a question and get an answer built only from your own documents, with deterministic citations and a confidence score. If the sources don't support an answer, it says so instead of inventing one.
Real processing pipeline
Text extraction, cleaning, and chunking run in the background. Embedding failures don't block a document from being usable.
Fail-closed auth
Organisation-scoped access where an auth failure denies rather than permits, with a regression suite covering the boundary.
Why it exists, and how it’s built.
What it’s for
Businesses accumulate contracts, policies, invoices, and reports faster than anyone can read them. The answer someone needs usually exists — it’s just buried in a file nobody remembers the name of.
This makes that pile searchable by meaning, then answerable in plain language, with every claim traceable back to the document and passage it came from. The citation part matters most: an AI answer about a contract is worthless if you can’t verify it.
How it’s built
A monorepo — Next.js 15 frontend, FastAPI backend, Supabase Postgres with pgvector for embeddings, Supabase Storage for files. Documents flow through extraction, chunking, and embedding, then retrieval feeds a grounded prompt with a minimum-score filter so weak matches never reach the model.
Status
In development. The backend is the mature part: storage, the processing pipeline, embeddings, semantic search, and the grounded answer engine are all built and validated against live data. Current work is the document synthesis layer and bringing the frontend up to match — today’s chat and search pages are validation tools rather than finished product surfaces.
Other things we’ve built.
ReviewWard
LiveAutomated review requests for local businesses that never gate a bad rating.
Learn more
UNIVRSLL
In developmentAI document intelligence: read a stack of documents at a glance, then ask questions with answers traced to the page.
Learn more
Clipreel
LiveTurns hours of footage into a thirty-second montage, with every cut landing in a pause.
Learn moreWant something like this, built for you?
Tell us what you have in mind, even if you’re not sure how it would work. The first conversation is free.