discovering.app
Plato converts experimental inputs into literature-grounded research end-to-end: idea generation, methodology design, executable analysis, LaTeX manuscript writing, domain-routed multi-source retrieval (arXiv, PubMed, Europe PMC, OpenAlex, Crossref, Semantic Scholar), citation validation, reviewer-panel revision loops, reproducibility manifests, and LangFuse observability — dashboard at discovering.app (Next.js 15 + FastAPI SSE backend).
Scientific automation usually stops at drafting prose or summarizing papers. Real research workflows need literature retrieval, method design, executable analysis, citation validation, reviewer critique, revision loops, and reproducibility records before the output can be trusted.
Plato structures the research pipeline into specialist stages that move from idea to method, results, manuscript, reviewer feedback, and reproducibility manifest. Retrieval is routed across scholarly sources while the dashboard streams reasoning and progress from the backend.
The difficult parts are grounding claims to citations, avoiding hallucinated references, preserving intermediate evidence, coordinating multiple agent stages, and making generated analysis reproducible enough for a human scientist to audit.
The system frames an AI scientist as an observable workflow, not a single black-box model. It pairs multi-source retrieval with citation validation, reviewer-panel loops, and stage-specific tools that produce a paper-like artifact with traceable evidence.
This demonstrates Stefan's domain fluency in AI research systems, scientific RAG, autonomous workflow design, FastAPI streaming, LangGraph-style orchestration, scholarly source routing, and reproducibility expectations.
Plato is strongest when read as research operations infrastructure. It turns literature search, method design, executable analysis, citation validation, reviewer critique, and reproducibility into a staged system rather than a single manuscript prompt.
The workflow has to track how claims originate, which sources support them, whether citations resolve, and how reviewer feedback changes the manuscript. That is the difference between a paper generator and an auditable research system.
Different scholarly sources answer different questions. Routing across arXiv, PubMed, Europe PMC, OpenAlex, Crossref, and Semantic Scholar lets the system separate prior work, biomedical evidence, citation metadata, and related-paper discovery.
The reviewer panel gives the system a way to criticize novelty, methods, statistics, and writing before final output. That loop matters because scientific output needs adversarial pressure before it becomes credible.
A serious scientific agent needs manifests, model versions, source ids, cost records, and evidence matrices. Those artifacts make the generated paper inspectable after the run, not just readable at the end.
LangGraph orchestration, retrieval adapters, and scientific executors share one typed Python boundary.
Paper pipeline nodes (idea, method, results, paper) compose as explicit graphs with checkpoint/resume.
Plato Dashboard frontend at discovering.app — stage cards, agent log streams, and cost tracking over SSE.
Dashboard gateway runs each Plato stage in subprocess workers and streams reasoning tokens to the UI.
End-to-end pipeline from data description to journal-styled LaTeX manuscript.
Next.js 15 + FastAPI SSE workspace at discovering.app with projects, models, costs, and activity routes.
Domain-routed adapters (arXiv, PubMed, Europe PMC, OpenAlex, Crossref, Semantic Scholar) with backoff and circuit breakers.
Resolves every reference against Crossref + Retraction Watch; emits validation_report.json.
Links atomic claims to source quote spans — evidence_matrix.jsonl per run.
Methodology, statistics, novelty, writing axes drive bounded redrafts.
Modal GPU, E2B sandbox, or local Jupyter run analysis per domain profile.
manifest.json with git sha, model versions, source ids, tokens, and cost.
plato loop iterates under wall-clock and cost budgets with regression revert.