AltheonShowcaseCase 02

Artificial Intelligence · Research Platform

AI-assisted research workflow platform for an academic institution.

A research department needed to reduce the time researchers spent on citation management, literature review, and knowledge organisation — and replace a fragmented toolset with a single, AI-integrated platform.

60%Reduction in research time
3 wksFrom spec to beta
40+Researchers onboarded
<200msAverage query response

The problem

Researchers were spending more time managing their tools than doing research. Citations were tracked in spreadsheets. Notes lived in disconnected documents. Literature searches required manually cross-referencing four or five different databases. There was no shared knowledge layer — when a researcher left the team, their institutional knowledge left with them.

The department had evaluated existing platforms like Zotero and Mendeley, but neither could handle their specific workflow requirements, particularly around AI-assisted synthesis and cross-project knowledge reuse.

What we built

We designed a research workflow platform with three core components:

  • Citation engine: Automatic citation extraction from PDFs, DOIs, and URLs, with export to all major formats (APA, MLA, Chicago, Vancouver) and bulk bibliography generation.
  • Knowledge graph: A vector-indexed library of all imported papers, with semantic search that surfaces relevant work across the full institutional corpus — not just a single researcher's library.
  • AI research assistant: A RAG-based assistant grounded entirely in the institution's own paper library. Researchers can ask questions, request summaries, and identify gaps in existing literature — with every answer sourced and citable.

"For the first time, a new researcher can get up to speed on three years of departmental work in a single afternoon. The institutional memory problem is solved." — Department Lead

AI architecture decisions

The assistant is deliberately closed — it only draws on the institution's own uploaded materials, not the open internet. This was a deliberate design choice: researchers needed answers they could cite and verify, not plausible-sounding summaries from unknown sources.

Documents are chunked, embedded via OpenAI's embedding API, and stored in Pinecone. At query time, the top-k chunks are retrieved and passed to GPT-4 with strict instructions to cite the source document and page for every factual claim. The average query-to-response time is under 200ms.

Outcome

The department reported a 60% reduction in time spent on literature review and citation management in the first semester of use. Forty researchers were onboarded within the first three weeks of launch, with adoption driven entirely by word of mouth within the department.

More Work