Manual record entry — structured fields feed straight into the matching pipeline
What's happening: a clinician creates a patient record and enters diagnosis, biomarker, and treatment fields one at a time. This structured profile is what the AI searches against. No free text required.
Patients
+ New patient
Diagnosis
Biomarkers
Treatment history
Matched evidence
Builds a query from the profile above and runs it through the RAG pipeline.
What's happening: the same patient could instead be described in plain language. Free text and structured intake feed the identical pipeline underneath.
Conversation
What Paxis does that manual review can't
The reasoning that ran underneath the answer you just saw
What's happening: nothing to click here — this is a plain-language look at the four mechanisms behind that last answer, so it's clear this wasn't a keyword search.
Implicit → explicit inference
Reads clinical language the way literature phrases it
The chart says one thing, the literature says another. Paxis's ontology-inference layer bridges them automatically:
"no longer a surgical candidate"→unresectableinoperable
"progressing on pembrolizumab"→ICI-refractoryanti-PD1 failure
Multi-axis parallel search
Searches 8 clinical axes at once instead of running one query
Cancer site, histology, biomarkers, treatment trajectory, comorbidities, and staging are each embedded and searched separately across Qdrant, PostgreSQL structured records, and patient-outcome frames in parallel, then merged. A single search box can't hold that.
Literature volume
More papers than one physician can screen
This patient's cohort alone has ~49 relevant studies published in the last few years. Screening that manually, on top of a full clinic load, isn't realistic every time a treatment decision comes up.
Cross-encoder relevance gate
Every candidate is re-scored against the full profile
A study matching on one keyword is not enough. Each candidate is re-ranked against the complete patient profile before it's ever shown, rejecting superficial matches automatically.
Patient Matching
Trials ranked against the active patient's full clinical profile
What's happening: each study below was independently scored 0–100% against this patient's profile. The bottom trial is excluded. The patient is no longer a surgical candidate, so it cannot apply.
Treatment Comparison
Trastuzumab deruxtecan vs. T-DM1 — extracted head-to-head from trial data
What's happening: efficacy and safety endpoints are pulled directly out of the trial publications and lined up side by side, so two regimens can be compared without reading both papers in full.
Case Alerts
Saved cases can watch the literature and notify you when something new matches
What's happening: alerts are already on for this saved case. Paxis keeps checking new literature in the background. Watch the notification appear when a new matching study is published.
Second-line salvage therapy in ICI-refractory HNSCC
Saved case · 80M, recurrent oral tongue SCC
This runs on its own. Nobody has to re-run the search.
Literature Analytics
Aggregated across 49 matched studies
What's happening: these numbers are computed live across every study matched to this patient's profile. Nothing here is a stored report.
Publications by year
My Collections
Auto-seeded and manually saved evidence for the active patient
What's happening: "Auto-seeded" studies were added by the matching pipeline automatically; "Saved by you" studies were added manually. Both stay organized per patient.
Trial Finder
Search open trials directly, independent of a specific patient
What's happening: this search runs against the open trial registry directly, so you can browse what is recruiting before any patient profile exists.
Dr. Demo Physician