A1live
Platform Core
Unified Postgres + Auth + RLS + Storage + consent + data export/delete.
Supabase · row-level security · AES-256 at restThe Zenlo engine
Zenlo isn't one app — it's a single clinical engine decomposed into atomic, reusable blocks across seven families. Each product is an assembly of the same blocks. This is the whole system, with an honest status on every part.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
A PDF, a photo, or raw text arrives. The Lab Test Analyzer extracts every value with OCR and turns it into structured biomarker JSON.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
Units are reconciled and each value is mapped to the Biomarker Library — 102 biomarkers with reference ranges by age and sex.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
The Clinical Pattern Engine tests the profile against 15 clinical patterns and 8 interaction types between them.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
HOMA-IR and the Biological Age Score are computed deterministically from the normalized panel.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
Repeat panels are compared with directional meaning — lower, higher, or U-shape is better — per biomarker.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
Every reading is checked against the Knowledge Store — 2,534 documents of population data and literature.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
A structured draft is written: composite, patterns, findings that need action, and trends.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
How the engine reads a patient
Eight steps, from an uploaded file to a structured clinical reading. Tap a step to see what happens inside it.
The draft lands in the product surface — a doctor dashboard, a patient portal, or an aggregate-only employer view.
One block, many products. The Clinical Engine (B1–B6) is reused across seven product configurations — the same parsing, the same patterns, the same biological-age and insulin-resistance math. Build a capability once; ship it everywhere. That reuse is the leverage behind a one-engineer company.
The 54 blocks · 7 families
Every atomic capability, its current status, and the AI model behind it where one is used. Filter by maturity — we label what's live and what's still on the bench.
The data, identity, and access layer everything else is built on.
Unified Postgres + Auth + RLS + Storage + consent + data export/delete.
Supabase · row-level security · AES-256 at restOne user gets revocable read / limited-write access to another's data.
RLS relations · patient / caregiver scopesRestricts who can create accounts on a given instance.
DB trigger · invite-gated cohortsThe shared brain. Reads biomarkers, finds patterns, quantifies risk — the most reused cluster in the system.
Reference catalogue of 102 biomarkers: code, name, unit, reference range by age/sex, category.
102 rules · unit normalizationPDF / photo / text / manual entry becomes structured biomarker JSON. This is the OCR + extraction step.
Claude Haiku 4.5 · ZDR · ~$0.001 / patient15 validated patterns plus interaction analysis across 8 relationship types between them.
Claude Haiku 4.5 · F1 0.963 best-in-benchmarkInsulin-resistance index from fasting glucose × insulin, with a clinical threshold.
deterministic · in-pipelineWeighted Z-score model across 8 biomarker categories vs age/sex reference ranges.
Claude Haiku 4.5 for the narrativeLongitudinal tracking with directional meaning (lower / higher / U-shape is better).
per-biomarker trend semanticsThe same dataset through five models with F1 comparison — the basis for our benchmark and preprints.
5 models · medRxiv + JAMIA OpenThe surfaces people actually touch — dashboards, chat, journals, document vaults.
A 4-tab experience (Stories / Numbers / Ask / You) — "Duolingo for your body".
Haiku 4.5 + Sonnet 4.6Full recovery shell: Today / Treatment / Meds / Labs / Team / Journal / Ask / Docs.
Haiku 4.5 + Sonnet 4.6Educational story templates that explain a result, with a safety post-filter.
Claude Haiku 4.5 + word filter4-layer memory: standing history + live context + retrieval + tools (trends, adherence, documents, schedule).
Sonnet 4.6 · Extended Thinking on serious casesMulti-type medical documents: labs, imaging, consults, prescriptions, discharge.
encrypted storage bucketSupplement / prescription banks + schedule + intake log + % adherence.
schedule + intake modelDaily mood/fatigue/sleep ratings + chip symptoms + an NIH-validated assessment.
PRO-CTCAE instrumentDoctor cards with specialty, facility, schedule, primary flag, caregiver block.
live in RecoveryUpcoming cycles, visits, labs, and pre-visit prep materials.
cycle + visit modelA ladder-of-trust funnel that collects profile data in tiers, with inline follow-ups.
staged data collectionApple Health / Oura / Whoop / Google Fit / Garmin / Fitbit connectors + a daily insight.
Claude Haiku 4.5 for daily insightSide-by-side "you see / employer sees" + a HIPAA / BAA / ZDR / RLS technical fold-out.
trust UX componentAlternate rendering: raw numbers, 6-month delta, reference-position bar, model info.
same data, denser viewAggregate-only workforce view. By architecture, never shows an individual's health data.
Company → admins → roster → communications → launch → day-0 countdown.
6-step setupMetric cards + monthly insight + cost-avoided YTD + attention items.
Claude Haiku 4.5 for insight copy5-stage adoption funnel + completion + department breakdown (engagement only).
aggregated server-sidePopulation pattern trends + benchmark vs norm. Per-department health is never shown, regardless of size.
Claude Haiku 4.5 for observationsModeled estimate with confidence interval, value categories, and a claim / don't-claim matrix.
transparent assumptionsExecutive view + QoQ curve + compliance grid + multi-length PDF export.
aggregate-only on exportStatus / department / invited / last-signed-in only. Cannot surface activity by design.
scope-isolatedFunnel metrics + recommended nudge + tone presets + automation timeline.
Claude Haiku 4.5 for tone variantsSeat usage + next invoice + payment method + billing history + ACH/PO.
Stripe (configured)Compliance status cards + data-flow diagram + subprocessor table + downloadable docs.
dynamic compliance statek ≥ 10 enforced server-side on every aggregate query, with visible suppression. Raisable, never lowerable.
API-layer enforcementAn autonomous research pipeline that runs on public datasets only (NHANES, BRFSS, PubMed) — fully separate from patient data and the production app.
2,534 documents in a vector store (1,349-doc bootstrap core) spanning population data and literature, 2000–2025.
pgvector · HNSW · OpenAI embeddingsWeekly PubMed saturation scan → a gap map by topic × year → top-3 topics.
Haiku 4.5 + Gemini for scoring7 hypothesis types per topic (confirmatory, contradictory, null, mediator-shift, subgroup-reversal, threshold-break, interaction-only), scored to top-3.
Sonnet 4.6 generates · Haiku scoresParallel testing: logistic regression (OR + 95% CI), interactions, stratified, threshold, propensity matching, survey weights.
Python · statsmodels · scipyStructured report: winning hypothesis, finding, product implication, confidence, rejected hypotheses, next step.
Claude Sonnet 4.65-model parallel review — structure, citations, novelty, adversarial critique, and a healthcare-tier clinical-validity model.
5 models · weighted approve / revise / rejectJournal-aware drafts with real citations pulled from the literature stage.
Claude Sonnet 4.6Post-processing dedup + a universal study-design validator across any topic.
Python utilitiesSafety layer: the agent reads broadly but writes only to a whitelisted workspace — no new pages, no deletions.
write-scope enforcementThe build system itself: how Zenlo ships code through AI agents under human review.
Command intake from dev / research / pm / marketing channels → the right agent group.
Node.js service · orchestrationPlanner → Generator → Evaluator with sprint contracts and automated browser testing.
Opus 4.7 plans · Sonnet 4.6 builds + grades6 lifecycle events — permission checks, auto-format, state-save before compaction.
pre/post tool · compaction hooksTimestamped branches + auto-commit + auto-push → preview deploy → human review.
git hooks + CIThe go-to-market and trust layer that wraps the technology.
Employer sourcing + LinkedIn + email + dialer + CRM pipeline.
connected SaaS toolingA ready pilot contract structure with a stop/go ROI gate.
e-sign readyA tiered per-member-per-month pricing model from pilot to enterprise.
published modelA structure for engaging practising MDs as clinical advisors.
sourcing in progressAI + healthcare-data consulting as parallel income during the build phase.
inbound via services listingsA claim / don't-claim matrix + a subprocessor registry + change-notification policy.
Anthropic BAA signed · ZDR activeA forbidden-word filter + tone registers per surface, regenerating copy when needed.
Claude Haiku 4.5 for regenerationThe 54 blocks · 7 families
Every atomic capability, its current status, and the AI model behind it where one is used. Filter by maturity — we label what's live and what's still on the bench.
The data, identity, and access layer everything else is built on.
Unified Postgres + Auth + RLS + Storage + consent + data export/delete.
Supabase · row-level security · AES-256 at restThe shared brain. Reads biomarkers, finds patterns, quantifies risk — the most reused cluster in the system.
Reference catalogue of 102 biomarkers: code, name, unit, reference range by age/sex, category.
102 rules · unit normalizationPDF / photo / text / manual entry becomes structured biomarker JSON. This is the OCR + extraction step.
Claude Haiku 4.5 · ZDR · ~$0.001 / patient15 validated patterns plus interaction analysis across 8 relationship types between them.
Claude Haiku 4.5 · F1 0.963 best-in-benchmarkInsulin-resistance index from fasting glucose × insulin, with a clinical threshold.
deterministic · in-pipelineWeighted Z-score model across 8 biomarker categories vs age/sex reference ranges.
Claude Haiku 4.5 for the narrativeLongitudinal tracking with directional meaning (lower / higher / U-shape is better).
per-biomarker trend semanticsThe same dataset through five models with F1 comparison — the basis for our benchmark and preprints.
5 models · medRxiv + JAMIA OpenThe surfaces people actually touch — dashboards, chat, journals, document vaults.
Doctor cards with specialty, facility, schedule, primary flag, caregiver block.
live in RecoveryAggregate-only workforce view. By architecture, never shows an individual's health data.
k ≥ 10 enforced server-side on every aggregate query, with visible suppression. Raisable, never lowerable.
API-layer enforcementAn autonomous research pipeline that runs on public datasets only (NHANES, BRFSS, PubMed) — fully separate from patient data and the production app.
2,534 documents in a vector store (1,349-doc bootstrap core) spanning population data and literature, 2000–2025.
pgvector · HNSW · OpenAI embeddingsWeekly PubMed saturation scan → a gap map by topic × year → top-3 topics.
Haiku 4.5 + Gemini for scoring7 hypothesis types per topic (confirmatory, contradictory, null, mediator-shift, subgroup-reversal, threshold-break, interaction-only), scored to top-3.
Sonnet 4.6 generates · Haiku scoresParallel testing: logistic regression (OR + 95% CI), interactions, stratified, threshold, propensity matching, survey weights.
Python · statsmodels · scipyStructured report: winning hypothesis, finding, product implication, confidence, rejected hypotheses, next step.
Claude Sonnet 4.65-model parallel review — structure, citations, novelty, adversarial critique, and a healthcare-tier clinical-validity model.
5 models · weighted approve / revise / rejectPost-processing dedup + a universal study-design validator across any topic.
Python utilitiesSafety layer: the agent reads broadly but writes only to a whitelisted workspace — no new pages, no deletions.
write-scope enforcementThe build system itself: how Zenlo ships code through AI agents under human review.
Command intake from dev / research / pm / marketing channels → the right agent group.
Node.js service · orchestration6 lifecycle events — permission checks, auto-format, state-save before compaction.
pre/post tool · compaction hooksTimestamped branches + auto-commit + auto-push → preview deploy → human review.
git hooks + CIThe go-to-market and trust layer that wraps the technology.
Employer sourcing + LinkedIn + email + dialer + CRM pipeline.
connected SaaS toolingA ready pilot contract structure with a stop/go ROI gate.
e-sign readyA tiered per-member-per-month pricing model from pilot to enterprise.
published modelAI + healthcare-data consulting as parallel income during the build phase.
inbound via services listingsA claim / don't-claim matrix + a subprocessor registry + change-notification policy.
Anthropic BAA signed · ZDR activeThe 54 blocks · 7 families
Every atomic capability, its current status, and the AI model behind it where one is used. Filter by maturity — we label what's live and what's still on the bench.
The data, identity, and access layer everything else is built on.
One user gets revocable read / limited-write access to another's data.
RLS relations · patient / caregiver scopesRestricts who can create accounts on a given instance.
DB trigger · invite-gated cohortsThe shared brain. Reads biomarkers, finds patterns, quantifies risk — the most reused cluster in the system.
No blocks at this maturity level.
The surfaces people actually touch — dashboards, chat, journals, document vaults.
Full recovery shell: Today / Treatment / Meds / Labs / Team / Journal / Ask / Docs.
Haiku 4.5 + Sonnet 4.6Multi-type medical documents: labs, imaging, consults, prescriptions, discharge.
encrypted storage bucketDaily mood/fatigue/sleep ratings + chip symptoms + an NIH-validated assessment.
PRO-CTCAE instrumentUpcoming cycles, visits, labs, and pre-visit prep materials.
cycle + visit modelAggregate-only workforce view. By architecture, never shows an individual's health data.
No blocks at this maturity level.
An autonomous research pipeline that runs on public datasets only (NHANES, BRFSS, PubMed) — fully separate from patient data and the production app.
Journal-aware drafts with real citations pulled from the literature stage.
Claude Sonnet 4.6The build system itself: how Zenlo ships code through AI agents under human review.
Planner → Generator → Evaluator with sprint contracts and automated browser testing.
Opus 4.7 plans · Sonnet 4.6 builds + gradesThe go-to-market and trust layer that wraps the technology.
A structure for engaging practising MDs as clinical advisors.
sourcing in progressThe 54 blocks · 7 families
Every atomic capability, its current status, and the AI model behind it where one is used. Filter by maturity — we label what's live and what's still on the bench.
The data, identity, and access layer everything else is built on.
No blocks at this maturity level.
The shared brain. Reads biomarkers, finds patterns, quantifies risk — the most reused cluster in the system.
No blocks at this maturity level.
The surfaces people actually touch — dashboards, chat, journals, document vaults.
A 4-tab experience (Stories / Numbers / Ask / You) — "Duolingo for your body".
Haiku 4.5 + Sonnet 4.6Educational story templates that explain a result, with a safety post-filter.
Claude Haiku 4.5 + word filter4-layer memory: standing history + live context + retrieval + tools (trends, adherence, documents, schedule).
Sonnet 4.6 · Extended Thinking on serious casesSupplement / prescription banks + schedule + intake log + % adherence.
schedule + intake modelA ladder-of-trust funnel that collects profile data in tiers, with inline follow-ups.
staged data collectionApple Health / Oura / Whoop / Google Fit / Garmin / Fitbit connectors + a daily insight.
Claude Haiku 4.5 for daily insightSide-by-side "you see / employer sees" + a HIPAA / BAA / ZDR / RLS technical fold-out.
trust UX componentAlternate rendering: raw numbers, 6-month delta, reference-position bar, model info.
same data, denser viewAggregate-only workforce view. By architecture, never shows an individual's health data.
Company → admins → roster → communications → launch → day-0 countdown.
6-step setupMetric cards + monthly insight + cost-avoided YTD + attention items.
Claude Haiku 4.5 for insight copy5-stage adoption funnel + completion + department breakdown (engagement only).
aggregated server-sidePopulation pattern trends + benchmark vs norm. Per-department health is never shown, regardless of size.
Claude Haiku 4.5 for observationsModeled estimate with confidence interval, value categories, and a claim / don't-claim matrix.
transparent assumptionsExecutive view + QoQ curve + compliance grid + multi-length PDF export.
aggregate-only on exportStatus / department / invited / last-signed-in only. Cannot surface activity by design.
scope-isolatedFunnel metrics + recommended nudge + tone presets + automation timeline.
Claude Haiku 4.5 for tone variantsSeat usage + next invoice + payment method + billing history + ACH/PO.
Stripe (configured)Compliance status cards + data-flow diagram + subprocessor table + downloadable docs.
dynamic compliance stateAn autonomous research pipeline that runs on public datasets only (NHANES, BRFSS, PubMed) — fully separate from patient data and the production app.
No blocks at this maturity level.
The build system itself: how Zenlo ships code through AI agents under human review.
No blocks at this maturity level.
The go-to-market and trust layer that wraps the technology.
A forbidden-word filter + tone registers per surface, regenerating copy when needed.
Claude Haiku 4.5 for regenerationThe 54 blocks · 7 families
Every atomic capability, its current status, and the AI model behind it where one is used. Filter by maturity — we label what's live and what's still on the bench.
The data, identity, and access layer everything else is built on.
No blocks at this maturity level.
The shared brain. Reads biomarkers, finds patterns, quantifies risk — the most reused cluster in the system.
No blocks at this maturity level.
The surfaces people actually touch — dashboards, chat, journals, document vaults.
No blocks at this maturity level.
Aggregate-only workforce view. By architecture, never shows an individual's health data.
No blocks at this maturity level.
An autonomous research pipeline that runs on public datasets only (NHANES, BRFSS, PubMed) — fully separate from patient data and the production app.
No blocks at this maturity level.
The build system itself: how Zenlo ships code through AI agents under human review.
No blocks at this maturity level.
The go-to-market and trust layer that wraps the technology.
No blocks at this maturity level.
The engine, packaged
A product is a target buyer plus a selection of blocks behind one surface. These are the configurations that are live or in build today.
AI early detection for self-insured workforces — aggregate risk, never individual health data.
A1 · B1–B7 · C1 · C3 · C4 · C10–C13 · D1–D11 · G-stackAI blood-test interpretation for functional-medicine practices — physician-ready reports the doctor signs.
A1 · A2 · B1–B6 · C4 · C5 · C13Post-treatment recovery & monitoring for oncology patients and their caregivers. Not diagnostic.
A1–A3 · B1–B6 · C2 · C4–C9More configurations — clinical API, research studio, caregiver, athlete, and others — are in design on the same core. See what's shipping first →