CopernicusAI · Gary Welz · CUNY Graduate Center
CopernicusAI is a working research prototype: scouts, embedding and a query interface run in production against three engines — GLMP (biology), ATAP (mathematics) and TDAP (topological data analysis, early stage). The architecture is domain-agnostic by design: what makes a field's engine is which literature its scouts watch and which flowcharts its researchers build, not a rewrite of the platform underneath. A new podcast briefing is generated only after at least three papers have been confirmed to exist in a registry (arXiv, PubMed or Crossref), and its reference list is built by code from those confirmed records. Quality scores and paradigm-shift notes are heuristic or model-generated, and are marked as such under Current Limitations.
The Research Tools Dashboard is a prototype web interface (first deployed December 2025) for exploring the knowledge graph, vector search, RAG question answering and content browsing. What is not yet verified or not yet built is listed plainly in the Current Limitations section.
CopernicusAI is organized in six tiers. A researcher works within a specific frontier — GLMP (biology), ATAP (mathematics) or TDAP (topological data analysis, early stage) — using the Knowledge Base UI as their instrument: the microscope they focus on that frontier, backed by Methods & Tools (the Programming Framework and its demonstration flowcharts). What they produce — podcasts, written papers — becomes new Resources, alongside metadata that Scouts bring in from the wider paper universe. A resource only becomes retrievable knowledge once it passes through Embedding, the gate between "exists in the collection" and "answerable by the Knowledge Base UI." That is the loop that makes this an engine rather than a pipeline: what a researcher produces today becomes what the next researcher finds tomorrow.
Figure: the researcher-workflow loop — what makes CopernicusAI an engine, not a pipeline.
Scouts run twice daily against PubMed and bioRxiv/medRxiv. arXiv acquisition is currently failing because the arXiv API is rate-limiting requests (see Current Limitations). A NASA ADS harvester exists in the repository but has contributed no papers. Scouts add metadata to Resources alongside products (podcasts, papers, videos) from the researcher's own work.
A resource becomes retrievable knowledge only once it is vectorized into the shared corpus (Firestore + GCS) with an OpenAI embedding model. Not stored — findable. Embedded records: loading….
Multiple access interfaces including RAG queries, vector search, knowledge graph visualization, API endpoints, and web interfaces, converging to unified knowledge output.
The headline diagram above is the researcher's loop. Internally, ingested material moves through three parallel modality lanes before converging at the same Embedding gate:
Figure: how processing works internally — three modality lanes converging at Embedding, then fanning out to the same Knowledge Base UI as the headline diagram.
Rather than one large index spread thin, the architecture is tested where it has been pointed at a real frontier. Each engine owns a set of declared research questions; its count below is the number of papers tagged to those questions, which is a smaller and sharper number than "papers in general". Counts are read live from the published status file.
Genome Logic Modeling Project: regulatory circuits modeled as logic-gate flowcharts, with a process-chart database.
Process charts: loading… · biology papers corpus-wide: loading…
Explore GLMP →Axiomatic Theories, Algorithms and Proofs: algorithms, axiomatic theories and proof structure modeled as dependency graphs.
Process charts: loading… · mathematics papers corpus-wide: loading…
Explore ATAP →Topological Data Analysis Project: a hosted engine built with an outside domain collaborator, on persistent cohomology and circular/toroidal coordinates. Early stage.
No process-chart database yet.
TDAP repository →
Engine counts come from the status file's papers_by_initiative and
process_databases fields. A paper's discipline is a different axis from the engine whose
questions it is tagged to, so the two numbers differ. Scouts tuned to each engine's research
frontier, rather than to its broad discipline, are in development.
CopernicusAI is a research prototype exploring AI-generated audio briefings as an interface for assisted scientific research. Listening is public at copernicusai.fyi.
Podcast generation is currently limited to allowlisted subscribers, each with a monthly quota. A request that cannot find enough confirmed sources fails with a notice rather than publishing unverified references.
Planned, not built: open generation for any user, iterative question refinement through a conversational interface, and collectively generated and shared research artifacts grounded in scientific sources.
The Research Tools Dashboard is a prototype deployed to Google Cloud Run (its page responded when last checked on 2026-10-20), offering knowledge-graph visualization, vector search, RAG queries and content browsing. It has not been through a formal evaluation.
See the "Knowledge Engine Ecosystem" section below for details.
Inspired by Nicolaus Copernicus's evidence-based challenge to accepted knowledge, CopernicusAI builds instruments for observing humanity's collective scientific knowledge — a growing index of research papers, regulatory and logical process flowcharts, and AI-generated podcast briefings in a single queryable knowledge graph. GLMP (biology) and ATAP (mathematics) are the two engines with process-chart databases; TDAP (topological data analysis) is a hosted engine at an early stage. A new briefing requires at least three papers confirmed to exist in a registry, and its references are built by code from those records; quality scores and paradigm-shift notes are heuristic or model-generated. The Copernican ethos applies to the platform itself: a marked limit is a finding, so limits are listed below.
An integrated ecosystem of research and collaboration tools designed to assist scientists in their workflow, from research discovery through knowledge synthesis to multi-format content generation. View Public Project Interface →
Synthesis & distribution platform for AI-powered research briefing podcast generation
Visit Website →Foundational meta-tool for universal process analysis across disciplines
Explore →Core data infrastructure for research paper metadata and citation networks
Explore →Prototype web interface for the knowledge graph, vector search, RAG queries and content browsing
Live prototype →Multi-voice AI podcasts of about 5–10 minutes that synthesize research papers. Generation is currently open to allowlisted subscribers only, each with a monthly quota. A subscriber can submit their episode to the public RSS feed; controls for private episode access are being rolled out in stages and are not yet complete.
What the production service calls today:
Vertex AI is disabled in production. No Anthropic model is called by the production service.
The corpus holds loading… papers, acquired through the scouts below. Sizes of the external databases are not claimed here.
Audio podcast system: generation for allowlisted subscribers and distribution through one public RSS feed, which Spotify, Apple Podcasts and YouTube read.
None of the following is built:
See: Science Video Database — companion project for research video content management.
A centralized metadata repository (not a file archive) of structured records. Scouts bring in metadata from PubMed and bioRxiv/medRxiv; embedding is what turns a stored record into retrievable knowledge.
Loading live counts from knowledge-engine-status.json…
Part of the GLMP research program — huggingface.co/spaces/garywelz/glmp
These platforms enable collective participation and collaboration across diverse user communities:
Like a microscope enables observation of the microscopic world, these tools enable observation and exploration of humanity's collective knowledge.
This platform represents prior work that demonstrates foundational research and development achievements in AI-powered scientific knowledge synthesis, collaborative research tools, and multi-modal content generation. These contributions establish the technical foundation and proof-of-concept for the broader CopernicusAI Knowledge Engine initiative.
This platform serves as the core synthesis and distribution component of the CopernicusAI Knowledge Engine. The Knowledge Engine is an integrated ecosystem of research and collaboration tools that work together to assist scientists in their workflow, from research discovery through knowledge synthesis to multi-format content generation.
The Knowledge Engine is designed to grow and evolve. Additional tools, databases, and collaboration components will be added as the project develops, expanding capabilities for AI-assisted scientific research and knowledge discovery.
For grant proposals:
Welz, G. (2024–2026). CopernicusAI: Knowledge Engine for Scientific Discovery.
Hugging Face Space. https://huggingface.co/spaces/garywelz/copernicusai
Live Platform: https://www.copernicusai.fyi
BibTeX Format:
@misc{welz2025copernicusai,
title={CopernicusAI: Knowledge Engine for Scientific Discovery},
author={Welz, Gary},
year={2024--2026},
url={https://huggingface.co/spaces/garywelz/copernicusai},
note={Hugging Face Space, Live Platform: https://www.copernicusai.fyi}
}https://copernicus-podcast-api-phzp4ie2sq-uc.a.run.app
Welz, G. (2024–2026). CopernicusAI: AI-Generated Audio Briefings as a Research Interface.
Hugging Face Spaces. https://huggingface.co/spaces/garywelz/copernicusai
BibTeX Format:
@misc{welz2025copernicusai,
title={CopernicusAI: AI-Generated Audio Briefings as a Research Interface},
author={Welz, Gary},
year={2024--2026},
url={https://huggingface.co/spaces/garywelz/copernicusai},
note={Hugging Face Space}
}NSF proposal pending.
The CopernicusAI Knowledge Engine is an integrated ecosystem of research and collaboration tools. The Research Tools Dashboard is a prototype (first deployed December 2025) with a web interface to the other components.
Web interface with knowledge graph visualization, vector search, RAG queries and content browsing over loading… papers.
Public Project Interface → (opens in new tab)Foundational meta-tool for universal process analysis across any discipline
Biology engine — loading… decoded processes in its process-chart database
Mathematics engine — algorithms, axiomatic theories and proofs as dependency graphs
Hosted engine, early stage — persistent cohomology and circular/toroidal coordinates
Core data infrastructure for structured research paper metadata and citation networks
Multi-modal content component with transcript-based search for scientific videos
Base URL: https://copernicus-podcast-api-phzp4ie2sq-uc.a.run.app
GET /api/content/browse?content_type=podcasts&limit=1
{
"content_type": "podcasts",
"items": [
{
"id": "ever-phys-250034",
"title": "3I/ATLAS: Unveiling the Secrets of Interstellar Comets and the Origins of Planetary Systems",
"type": "podcast",
"metadata": {
"duration": "5:32",
"category": "Physics",
"slug": "ever-phys-250034",
"episode_link": "https://copernicusai.fyi/episodes/ever-phys-250034"
}
}
],
"pagination": { "page": 1, "limit": 1, "total": 39, "pages": 39 }
}Recorded 2026-10-20; the total changes as episodes are published.
The endpoints listed here are read-only and need no credentials. Routes that write, generate podcasts or administer the service require credentials and are not documented here.
/api/rag/answer allows 20 requests per hour per client address. No general rate limit or API-key quota is published for the other endpoints.
The API is under active development and has no versioning or backward-compatibility guarantee. Machine-readable definitions are served at /openapi.json.