Find what's missing. Fill the gaps. Cite every hop.
The bias-free biomedical reasoning layer your charting tool can't be. One agent traversing the whole chain — symptom to mechanism to target to compound to feasibility to safety — with a primary-source citation at every step. Built for complex cases.
Cited at every hop · Bias-free, nothing to sell · Therapeutic reach computed · Not a medical device
Built on the evidence stack you already trust
All numbers from the gold-blind, fail-closed benchmark harness. Per-task n, metric, and reproduction commands in the methodology dossier.
Your tools answer one question at a time. Complex cases need the whole chain.
A symptom checker stops at the symptom. A guideline tool stops at the guideline. A literature database stops at the abstract. A drug-target graph stops at the association. Each of them answers one slice and hands the rest to you to stitch together by hand.
For a routine case, that's fine. For a complex one — the multi-system patient with “normal” labs, the rare-disease odyssey that's run six years and seventeen encounters, the desk-phase research question that spans target discovery, polypharmacology, and safety — the stitching is where the time goes and the misses happen.
Kitrus walks the whole chain in one cited pass. Same engine, bidirectional: up the chain from symptom to root cause, down from disease to candidate intervention. Every hop carries a primary-source citation. The feasibility computation that other tools don't run — does the labeled dose plausibly reach the target above its potency threshold? — is surfaced inside the reasoning, not as an afterthought.
One agent. The whole chain. Provenance at every hop.
Kitrus's substrate is a knowledge graph — 262 million evidence rows across 258 source databases, with 15.6 million directed, typed edges and 16.1 million provenance rows. On top of that substrate, an agent walks. The walk is bidirectional. The trace is the product.
- 1
Mechanism map — disease to causal genes and pathways, directed (not merely associated). Sourced from Open Targets, Monarch, HPO, MSigDB with per-edge PMIDs.
- 2
Candidate targets + direction — which targets, and which way to move them (agonise / inhibit). Drug-target inference scored against TDC DTI-DG at parity with the leaderboard.
- 3
Candidate matching — compounds that modulate those targets in the needed direction, from ChEMBL and BindingDB with binding-affinity detail.
- 4
Therapeutic-reach feasibility — an advisory gate computing whether the labeled dose plausibly reaches the target above its potency threshold. ~93% agreement on a known-medication validation set.
- 5
Safety screen — adverse-event, interaction, and depletion filtering from FAERS, openFDA, and FDA labels.
- 6
Steelman-against — an adversarial pass that argues each surviving candidate down before it is shown to the user.
A benchmark per joint of the chain. Independently scored. Fail-closed.
Each link in the chain is independently benchmark-verified, version-pinned, and scored gold-blind by an official harness that fails closed — it abstains rather than coercing a gold answer. A chain is only as trustworthy as its weakest link; the suite below is a unit-test map across the reasoning chain.
Some joints land at substrate-native dominance. Some land at parity with specialist tools we absorb as steps. Some honest cedes — the wet-lab wall and contamination-prone leaderboards — we mark explicitly. The full per-task n, metric, variance, and one-line reproduction commands are in the methodology dossier.
HPO phenotype→gene packet
Gold causes_disease edges + LoF-mechanism flag
ClinVar 1.00 · CAGI/ClinGen VUS + ACMG PVS1 fires
Mechanism + protein-function packets · directed neighborhood traversal
Disease packet + MSigDB / GTRD / P-HIPSTer membership (1.00 / 0.975)
TDC DTI-DG 0.591 ≈ leaderboard 0.588
ChEMBL / BindingDB affinity
Safety facade · HealthBench drug-safety +0.127
What we cede: trained-predictor leaderboards (parity or below — we wrap the best specialist rather than out-build it), contamination-prone KG link-prediction, and the wet-lab wall. Honest scope boundaries, not gaps.
The find-and-fill-in cockpit for complex patients.
The hardest cases are multi-system, chronic, and not protocol-shaped: metabolic / hormonal patients with “normal” labs and unresolved symptoms. Existing tools answer one question at a time — drug interactions in one place, depletion lookups in another, guideline search in a third — and there is no bias-free tool that reasons across the seams.
Kitrus's clinical-operator cockpit gives the clinician one screen that holds the whole picture: the open hypotheses, the labs and medications that support or weaken them, the missing data that would most reduce uncertainty, and the next step that closes each gap. Every claim cites its primary source.
The cockpit is deployed inside a licensed partner's practice — UK CQC / GMC / GPhC, US MSO + physician-owned PC — so the clinical decision and the prescribing licence stay with the clinician, not the software. The artifact the clinician acts on is the Doctor Brief: a structured patient summary the clinician reviews, edits, and signs off, never a re-keyed intake.
WHERE THIS IS DEPLOYING — Australia (AU clinics, advisor-led) · United Kingdom (telehealth + pharmacy partner) · United States (MSO + physician-owned PC template)
The desk-phase walk that used to take a team takes one cited pass.
The in-silico stack is normally siloed — target discovery, polypharmacology, ADMET / PK, adverse-event safety, mechanism, literature evidence — each in its own tool, data model, and team. Moving between them is a manual hand-off, context is lost at every seam, and “for this disease, what is a feasible, safe candidate and why” takes weeks of literature-and-database assembly.
Kitrus runs the desk phase as one auditable, cited pass: mechanism map → candidate targets and direction → polypharmacology matching → therapeutic-reach feasibility → safety screen → steelman-against. Every hop is provenance-cited (PubMed, trials, drug labels). The feasibility computation flags whether the labeled dose plausibly reaches the target above its potency threshold — surfaced to the searching agent as an advisory signal, never used to silently discard a candidate.
This is the desk phase up to the wet-lab wall — coverage, not a moat everywhere. We rank where to spend wet-lab budget; we do not run assays. Weeks of manual triage become a cited, feasibility-screened shortlist.
Lean teams self-serve from $99/mo BYO-AI. Discovery pilots, team licences, business-unit deals, and enterprise are scoped on the call.
ACMG inside the cross-chain context.
Variant interpretation is labor-intensive: ClinGen credits up to 6 hours per approved curation; a hearing-loss VCEP study found 40 minutes average per variant; genetic counselors cite time as a top barrier. Existing classifiers — VarSome, Franklin, Fabric — are mature and clinically validated, but they classify the variant. They do not interpret it in the context of the whole patient.
Kitrus augments rather than displaces. We classify ACMG, but we add the layer the classifiers don't have: the disease-mechanism reasoning above classification, the cross-chain context (what other genes, drugs, exposures, and depletions interact with this variant), and a patient-facing brief alongside the clinician summary. Cited per evidence line — PMIDs and ClinVar VCV IDs surfaced inline.
For a clinical-genetics lab, this is an expansion play: the variant classifier you already use, with a cited mechanism + patient-context layer that produces the report your geneticists currently assemble by hand.
The reasoning endpoint your agent calls mid-workflow.
The modern knowledge-work pattern is bring-your-own-AI. A researcher or clinician increasingly works inside their own agent — an enterprise copilot carrying the organisation's context, or their own ChatGPT / Claude / API setup with their own information bank — and reaches out from it to external sources for what the base model cannot do alone.
Kitrus is built to be one of those sources. A safe, gated external-access function exposes the cross-chain walk: your agent calls Kitrus mid-workflow, gets a cited reasoning trace back, and continues. Symptom → mechanism → candidate → feasibility, in one fluid step instead of stitching together a dozen single-domain lookups by hand. The CLI / facade surface that exposes this is already in production.
This is the high-margin engine-licensing motion. Priced against the expert labor and the failures it removes — not per seat against a chat subscription. Pilot → team licence → enterprise.
BYO-AI: you bring the LLM key + compute, we charge for the reasoning + substrate + governance layer. Three tiers from $99/mo (BYO-key required) to $1,499/mo (BYO-key or hosted with PAYG overage). PAYG: $1 per credit, $20 minimum, volume discounts above $500. Generative-design tools require verified-org KYC on Pro / Scale / PAYG; Starter is reasoning + retrieval + light compute only. Enterprise licence + Startup / Lab middle tier — talk to us.
Nothing to sell. Nothing to recommend. Nothing to bias the reasoning.
Kitrus's reasoning layer has no commercial loops attached. We don't sell labs, we don't sell supplements, we don't take affiliate commissions, we don't accept pharma sponsorship, and we don't insert sponsored results.
When telehealth fulfilment exists — and it will, for the categories where it makes sense (GLP-1, weight, hair-loss, sexual function, longevity) — it sits behind a structural firewall. The reasoning never ranks, steers, or filters by what Kitrus's telehealth pharmacy sells. The engine recommends the right action even when that action is “nothing,” “a product we do not carry,” or “discuss this with your own clinician.”
Held, the line lets telehealth fund the consumer flywheel without touching the trust that powers it. Blurred, the differentiator is gone. So the separation is a hard product rule, not a preference.
What Kitrus is not.
Not a medical device
Kitrus is an informational decision-support tool. It investigates, surfaces hypotheses, and prepares clinician discussion points. It never diagnoses, treats, or prescribes. The clinician carries every clinical decision.
Not a replacement for clinical judgment
The trace is the product. We show the evidence and the reasoning so the clinician can audit, override, or accept. Confidence is stated; uncertainty is named.
Not a storefront
The reasoning layer never recommends Kitrus products. Telehealth fulfilment, where it exists, is structurally separated. The engine recommends the right action — including “do nothing” or “see your own clinician.”
Pre-scale regulatory classification review (EU MDR / IVDR, EU AI Act, UK MHRA, US posture) is scheduled before paid scale. Explicit gate, not an afterthought.
Honest about both sides.
All-green grids don't survive diligence. Here's where Kitrus shines today, and where well-funded incumbents lead — and how we're playing each side.
Where Kitrus shines today
Where the engine is built, we win. Where go-to-market is not yet there, we say so. This is the line that makes the rest credible.
Priced against the expert labor we replace, not per chat seat.
The economics are buyer-specific because the value is buyer-specific. A pharma target-discovery group avoiding a $255M Phase III dead-end is not on the same value scale as a solo functional-medicine clinician saving 30 minutes per chart. We price each surface against the failure mode it removes.
Indicative ranges below. Exact pricing is conversation-dependent; contracts on request.
Bring your own LLM key
From $99/mo
Subscription metered in credits. Starter requires BYO-key (your OpenAI / Anthropic / OpenRouter + Modal) for pure-margin pricing. Pro and Scale: BYO-key or hosted with PAYG overage. Generative-design tools gated behind verified-org KYC on Pro+. 3 tiers from $99/mo to $1,499/mo.
Use the engine, buy credits
From $1/credit, $20 min
Pay-as-you-go. Kitrus provides the LLM + engine. Credits proportional to compute (~1 credit / reasoning query, ~2 / fold, ~150 / generative campaign). No subscription. Volume discounts above $500. Generative-design requires KYC.
Recurring middle tier for lean biotechs
Talk to us
The bridge between Scale and Team Licence. Annual contract, BYO-compute orchestration option, KYC for design tools, modest CSM hours, 30-day pilot included. For organisations that outgrow self-serve Scale but aren't at full Team Licence yet.
Annual licence for organisations
Talk to us
Team Licence, Business Unit, or BYO-AI Endpoint Team Licence. Multi-user, BYO-compute option, dedicated CSM. Priced against the expert labor and failures the engine removes.
SLAs, custom integration, dedicated support
Talk to us
Multi-BU pharma, FHIR / EHR / LIMS integration, 99.9% uptime SLA, dedicated solutions engineer, private deployment option.
Fixed-fee one-off engagements
Talk to us
Disease Mechanism Map, Repurposing / Target Discovery Sprint, Program Red-Team Diligence, Discovery OS Design Partner. Deliverable-based, scoped on the call.
Workflow-scoped pilot (front door)
Talk to us
6–8 weeks, measured before / after. Success criteria you define with us at kickoff. Refundable in full if criteria aren't met. Converts to subscription or licence.
Self-serve tiers (BYO-AI + PAYG) are metered endpoint access + light compute. Contracted tiers add private-data integration, BYO-compute orchestration, the clinical cockpit, SLAs, and dedicated support — categories self-serve does not cover.
If the pilot's success criteria aren't met at week 6, the pilot is on us.
Discovery pilots are scoped against measurable before / after criteria you define with us at kickoff. If those criteria aren't met at week 6 — no questions, no clawback paperwork. The pilot is refunded in full.
Who Kitrus isn't for.
We disqualify aggressively because misfits hurt us both.
- Routine cases that a single-question guideline tool handles fine — you don't need cross-chain reasoning for ankle sprain triage.
- Buyers who need a fully autonomous prescription engine — Kitrus surfaces evidence; the clinician makes every decision.
- Teams without a problem they're paying expert labor to solve today — our pricing is anchored against the specialist hours we replace; if that hour isn't being spent, the value isn't there.
- Organisations that require EHR-native integration on day one — we integrate via clinical partner workflows and API today; native EHR surfaces are on the roadmap, not shipped.
- Buyers expecting clinical-trial-validated AI outputs in the variant diagnostic space — our ACMG classification is research-grade, third-party validation is in progress.
- Anyone looking for a chatbot that sells supplements — that is the anti-Kitrus.
Questions buyers actually ask.
Richard Oak
Medical advisor (AU)
Owns and operates two clinics; navigates AU / TGA regulation for the clinical wedge.
Theodore Agranat
Advisor
Serial technology entrepreneur and investor; founder of multiple web / marketing-technology startups since 1998.
Vihan Patel
Advisor
UK telehealth and pharmacy connections; opens doors for the clinical loop.
See the cited trace on a real case in 30 minutes.
We'll bring a sample case from your domain (or use yours, with synthetic data) and walk the cross-chain reasoning trace end-to-end. You'll see the evidence, the feasibility flag, the safety screen, and the discussion points the cockpit would surface.
If there's mutual fit at the end of 30 minutes, we'll scope a pilot. If not, you walk away with the methodology dossier and a clearer view of the layer.
30 minutes · No sales pitch · Cited methodology PDF on the way out