Ourobion
Agentic research system · One Health app · ASEAN
Health apps collect well and explain badly. Ourobion reads the scientific literature for you — then makes a second AI, from a different company, try to prove the finding wrong before it ever reaches a person.
Non-diagnostic by construction. Not a medical device.
The problem
Ourobion began with signals most apps overlook — gut and urinary health, hydration — and a practical limit. A health company hires people to find patterns and read the research behind them. We are three builders; unable to hire that department, we grew one.
The obvious fix — attach a language model, ask it why — is the most dangerous thing you can put in front of a health claim. A model will state a correlation as a cause, reverse which came first, and stretch a mouse study onto a person.
The problem is not data or model access. It is grounding the interpretation, when the best tool for it is the least trustworthy part of the system.
Before building, we fixed one criterion: no research claim reaches a person unless it points back to its paper and survives review by a model that did not write it.
The brain
James Reason’s Swiss-cheese model describes how harm reaches the end of a complex system — only when the gaps in successive layers happen to align. The goal is never a perfect layer. It is layers that fail differently.
A model can hallucinate a quotation that reads perfectly. A string comparison does not care how plausible it reads — it simply reports that those words are not in that paper. Code cannot be talked into agreeing; a model catches meaning no rule anticipates. Interleaving them means an error must survive two unrelated kinds of scrutiny.
Two models from one company share training data, methods, and therefore blind spots — so
asking one to check the other is theatre. The drafting model (OpenAI gpt-5) and
the reviewing model (Agnes AI agnes-2.5-flash) must come from different vendors.
The router refuses to start otherwise. It fails closed.
Any layer can stop the run. A supporting verdict additionally requires the reviewer to have retrieved its own evidence — without fresh grounding, the strongest verdict available to it is uncertain.
Evidence
Snapshot, not a live feed. Measured 3 August 2026. This page is static and reads no database — these figures were correct at that date and revision, and nowhere else. Current counts require re-running the pipeline.
Viceroy judges whether a sentence claims a cause or only a correlation. It scored 0.866 where a keyword-matching baseline scored 0.507 (macro-F1).
Zebra judges whether evidence supports a claim. It scored 0.599 against a bar of 0.70 written down before training — it failed, and the code refuses to promote it.
Both disagreed substantially with a third model on 96 real papers. With no adjudication, that measures disagreement, not correctness.
In nao, any one shows the drafted claim, the quoted spans it rests on, the citation, the reviewer’s verdict and confidence, and the decision that followed.
“Verified” has a narrow meaning here: the claim is a faithful reading of the paper it cites. It does not mean the wider literature agrees.
Two surfaces, one brain
The phone app — Flutter, Android
Thirty-second daily logging of signals the category skips: digestion, hydration, energy. Then descriptive insight cards about patterns in your own data — observational language only, never a diagnosis.
Debug-signed sideload build for Android. Not a Play Store artifact.
The research workbench — Next.js on Cloudflare
Where the pipeline is inspectable: the corpus, each proposed relationship, the evidence behind it, the independent reviewer’s verdict, and whether it was served, qualified, or withheld.
Sign in with the reviewer credentials below — viewer access, read-only.
Reviewer access — test@ourobion.com / test123, one credential, view-only
in both. In biotope it opens a seeded demonstration profile whose health logs are
synthetic; the cards and the relationships behind them are real, the health data
is not.
Future work
A card can currently be served on the strength of a single paper. The gate asks whether a claim is faithful to the paper it cites; it does not yet ask whether the field agrees. Closing that gap is the roadmap. Four layers are planned, ordered by how soon they are reachable — each one adds holes in a new place, which is the whole design.
Split papers into sections with stable sentence ids, so a quote is pinned to this sentence — making a stitched quotation detectable, not merely unlikely.
Follow a paper’s own references to place a claim in a literature. Deterministic graph traversal — the substrate corroboration needs.
When several papers independently support the same relationship, that agreement becomes a reliability signal. Not built: the model-call volume exceeds our budget.
Recombine the ranking signals into one interpretable score that can withhold again — without the failure mode that forced the change.
The goal these build toward: a served relationship that is not merely faithful to one paper, but corroborated across several — carrying a reliability score a person can read, and a gate that can withhold on the strength of the wider literature rather than a single citation.
Built by three