nspiration Janell Green Smith. Janell was a vibrant, healthy Black mother whose preventable pregnancy-related death is one of the thousands behind a sobering CDC statistic: Black women in the U.S. die from pregnancy-related causes at three times the rate of white women, and the majority of those deaths are preventable. Too often, the warning signs are there in the chart, in the vitals, in what the patient is telling the nurse — but they get missed in the rush of a busy floor, lost between EHR clicks, or never make it from intake to the attending in time.
GreenCompass is built in Janell's name. Our north star: catch what shouldn't be missed, and make sure no mother is sent home with a red flag the system failed to surface.
What it does GreenCompass is a clinical decision-support layer that sits next to Epic and accompanies the bedside nurse through OB intake (prenatal, labor & delivery, postpartum) and gynecologic concerns. In about 90 seconds it:
Pulls the chart from Epic — gestational age, gravida/para, blood type, allergies, active problems, recent BP and weight pre-fill automatically, so the nurse only answers what isn't already known. Walks the nurse through a focused, evidence-based questionnaire built around real red flags: severe-range BP, preeclampsia warning signs, hemorrhage, ruptured membranes, decreased fetal movement, sepsis, VTE, intrapartum eclampsia, uterine rupture, postpartum endometritis, and more. Triages the case in real time — every answer feeds a flag engine that classifies the encounter as green / amber / red, with explicit "escalate now" guidance for any red-flag combination (e.g. severe-range BP + headache → severe pre-eclampsia pathway). Drafts an AI clinical note in SOAP format using the answers and chart facts, ready for the nurse or attending to review before filing. Closes the loop with patient-matched support — based on visit type and risk level, it surfaces vetted communities like Black Mamas Matter Alliance, Postpartum Support International, and the Preeclampsia Foundation so women leave with peer support, not just a discharge sheet. How we built it Frontend: React 18 + Vite + TypeScript + Tailwind, with a hand-built design system (semantic tokens, GreenCompass green primary, editorial Fraunces + Inter pairing) so the clinical UI feels calm and unmistakably ours rather than generic SaaS. Backend: Lovable Cloud (Supabase) for auth, Postgres with row-level security on every table, and edge functions for the AI note generation and NPI provider lookup. Provider access to mock Epic charts is gated by a verified flag on provider_profiles that users cannot self-set — locked down with an RLS policy and a trigger that blocks self-elevation. AI: Lovable AI Gateway (Gemini 2.5) for SOAP-note drafting from structured intake answers + chart facts, with strict prompts that keep the model in clinician-supporting territory ("draft, review before filing"). Clinical logic: A typed pathway engine (pathways.ts, flag.ts, nextSteps.ts, summary.ts) where each visit type is a list of questions, and a deterministic rules engine — not the LLM — decides red/amber/green and the next-steps checklist. The AI never makes the triage call. Mock Epic integration: A mock_epic_patients table mirrors realistic FHIR-style chart facts so we can demo chart auto-fill end-to-end without a live Epic sandbox. Showcase video: A 60-second Remotion-rendered MP4 walks through the story for clinician and hospital-buyer pitches. Challenges we ran into Getting clinical scope right. It's tempting to ask everything; the real skill was cutting until every question was load-bearing for triage. We rewrote the prenatal and labor pathways multiple times to add the questions a nurse actually needs (intrapartum preeclampsia, sudden tearing pain → uterine rupture, dyspnea/SpO₂ → AFE/PE) and remove anything that didn't change a decision. Security around mock chart data. Our first cut let any signed-in user be treated as a "provider" — which would've meant any account could read all patient charts. We hardened it with a non-user-settable verified column, a trigger that prevents self-verification, and removed the self-insert policy entirely. AI in a clinical context. We had to draw a hard line: the LLM drafts notes, it does not triage. Flag logic is fully deterministic and testable. Prompts are constrained, outputs are framed as drafts requiring review. Rendering a 2-minute Remotion video in a sandboxed environment with a 600s render budget — we settled on a tight 60s cut at 1080p that still tells the full story for a clinician-buyer audience. Designing for trust. Nurses and OB attendings have zero patience for cluttered, "AI-vibes" UI. Getting the typography, spacing, and color signaling (red/amber/green pills) to feel like clinical equipment instead of a startup landing page took several passes. Accomplishments that we're proud of A real, end-to-end clinical workflow: chart auto-fill → focused intake → deterministic red-flag triage → AI SOAP note → matched patient support communities — not a mockup. A pathway library covering prenatal, labor & delivery, postpartum, and gynecologic visits with red flags grounded in CDC, ACOG, and NICHD guidance (severe-range BP, NICHD FHR categories, PALM-COEIN, Amsel criteria, CDC PID minimums). A genuinely opinionated visual identity — editorial typography, warm cream + evergreen palette, custom compass mark — that doesn't look like every other healthcare AI demo. A security posture we'd be comfortable taking into a real hospital conversation: RLS on every table, verified-provider gating, no PHI in client storage, no roles in profile tables. A 60-second showcase video, rendered programmatically with Remotion, that we can hand to any clinical buyer. Naming it for Janell — and making sure the product reflects her. What we learned Maternal mortality is a workflow problem as much as a clinical one. The data needed to save lives is usually already in the chart; what's missing is the right thing surfaced at the right second to the right person. The LLM is the assistant, not the clinician. Deterministic rules for triage, AI for the parts that are actually language work (drafting, summarizing, matching communities). That split made the product more trustworthy and easier to build. Equity has to be designed in, not bolted on. Matching patients to Black-led and culturally-specific support organizations (Black Mamas Matter Alliance, Sista Midwife Productions) is a product decision, not a marketing one. Security review must be continuous. Our first is_provider() check looked correct and was actually a privilege-escalation hole. Running the security scanner on every meaningful migration caught it. A real design system pays for itself fast. Once tokens were in place, every new pathway, card, and flag pill snapped into a consistent clinical look without one-off styling. What's next for GreenCompass Live Epic (FHIR) integration. Replace the mock chart with a real SMART-on-FHIR launch from inside Hyperspace, scoped to the active patient encounter. Bidirectional charting. Write the finalized SOAP note and the triage flag back into Epic as a structured progress note + flowsheet row, with the nurse's edits preserved. Closed-loop escalation. When a red flag fires, page the OB attending and rapid-response team directly from the app with a one-tap acknowledgment, and log time-to-acknowledgment for QI. Postpartum follow-up beyond discharge. Day-3, week-2, and 6-week check-ins via secure SMS for postpartum BP, mood (EPDS), bleeding, and warning signs — the period when most preventable deaths actually happen. Equity dashboard for hospital leadership. De-identified outcome metrics by race, payer, and language to make disparities visible at the unit and system level. Verified provider directory + warm referrals. Move from listing communities to actually scheduling a doula, lactation consultant, or perinatal mental health clinician before the patient leaves the floor. Pilot with a delivery hospital. Move from demo to a measured pilot on a single L&D unit and publish what we learn — including what doesn't work. In Janell's name. Catch what shouldn't be missed.
Here's a filled-out version tailored to GreenCompass — a clinician-facing OB triage assistant with Epic chart auto-fill, AI SOAP-note drafting, red-flag detection, and matched patient support communities.
Inspiration Janell Green Smith. Janell was a vibrant, healthy Black mother whose preventable pregnancy-related death is one of the thousands behind a sobering CDC statistic: Black women in the U.S. die from pregnancy-related causes at three times the rate of white women, and the majority of those deaths are preventable. Too often, the warning signs are there in the chart, in the vitals, in what the patient is telling the nurse — but they get missed in the rush of a busy floor, lost between EHR clicks, or never make it from intake to the attending in time.
GreenCompass is built in Janell's name. Our north star: catch what shouldn't be missed, and make sure no mother is sent home with a red flag the system failed to surface.
What it does GreenCompass is a clinical decision-support layer that sits next to Epic and accompanies the bedside nurse through OB intake (prenatal, labor & delivery, postpartum) and gynecologic concerns. In about 90 seconds it:
Pulls the chart from Epic — gestational age, gravida/para, blood type, allergies, active problems, recent BP and weight pre-fill automatically, so the nurse only answers what isn't already known. Walks the nurse through a focused, evidence-based questionnaire built around real red flags: severe-range BP, preeclampsia warning signs, hemorrhage, ruptured membranes, decreased fetal movement, sepsis, VTE, intrapartum eclampsia, uterine rupture, postpartum endometritis, and more. Triages the case in real time — every answer feeds a flag engine that classifies the encounter as green / amber / red, with explicit "escalate now" guidance for any red-flag combination (e.g. severe-range BP + headache → severe pre-eclampsia pathway). Drafts an AI clinical note in SOAP format using the answers and chart facts, ready for the nurse or attending to review before filing. Closes the loop with patient-matched support — based on visit type and risk level, it surfaces vetted communities like Black Mamas Matter Alliance, Postpartum Support International, and the Preeclampsia Foundation so women leave with peer support, not just a discharge sheet. How we built it Frontend: React 18 + Vite + TypeScript + Tailwind, with a hand-built design system (semantic tokens, GreenCompass green primary, editorial Fraunces + Inter pairing) so the clinical UI feels calm and unmistakably ours rather than generic SaaS. Backend: Lovable Cloud (Supabase) for auth, Postgres with row-level security on every table, and edge functions for the AI note generation and NPI provider lookup. Provider access to mock Epic charts is gated by a verified flag on provider_profiles that users cannot self-set — locked down with an RLS policy and a trigger that blocks self-elevation. AI: Lovable AI Gateway (Gemini 2.5) for SOAP-note drafting from structured intake answers + chart facts, with strict prompts that keep the model in clinician-supporting territory ("draft, review before filing"). Clinical logic: A typed pathway engine (pathways.ts, flag.ts, nextSteps.ts, summary.ts) where each visit type is a list of questions, and a deterministic rules engine — not the LLM — decides red/amber/green and the next-steps checklist. The AI never makes the triage call. Mock Epic integration: A mock_epic_patients table mirrors realistic FHIR-style chart facts so we can demo chart auto-fill end-to-end without a live Epic sandbox. Showcase video: A 60-second Remotion-rendered MP4 walks through the story for clinician and hospital-buyer pitches. Challenges we ran into Getting clinical scope right. It's tempting to ask everything; the real skill was cutting until every question was load-bearing for triage. We rewrote the prenatal and labor pathways multiple times to add the questions a nurse actually needs (intrapartum preeclampsia, sudden tearing pain → uterine rupture, dyspnea/SpO₂ → AFE/PE) and remove anything that didn't change a decision. Security around mock chart data. Our first cut let any signed-in user be treated as a "provider" — which would've meant any account could read all patient charts. We hardened it with a non-user-settable verified column, a trigger that prevents self-verification, and removed the self-insert policy entirely. AI in a clinical context. We had to draw a hard line: the LLM drafts notes, it does not triage. Flag logic is fully deterministic and testable. Prompts are constrained, outputs are framed as drafts requiring review. Rendering a 2-minute Remotion video in a sandboxed environment with a 600s render budget — we settled on a tight 60s cut at 1080p that still tells the full story for a clinician-buyer audience. Designing for trust. Nurses and OB attendings have zero patience for cluttered, "AI-vibes" UI. Getting the typography, spacing, and color signaling (red/amber/green pills) to feel like clinical equipment instead of a startup landing page took several passes. Accomplishments that we're proud of A real, end-to-end clinical workflow: chart auto-fill → focused intake → deterministic red-flag triage → AI SOAP note → matched patient support communities — not a mockup. A pathway library covering prenatal, labor & delivery, postpartum, and gynecologic visits with red flags grounded in CDC, ACOG, and NICHD guidance (severe-range BP, NICHD FHR categories, PALM-COEIN, Amsel criteria, CDC PID minimums). A genuinely opinionated visual identity — editorial typography, warm cream + evergreen palette, custom compass mark — that doesn't look like every other healthcare AI demo. A security posture we'd be comfortable taking into a real hospital conversation: RLS on every table, verified-provider gating, no PHI in client storage, no roles in profile tables. A 60-second showcase video, rendered programmatically with Remotion, that we can hand to any clinical buyer. Naming it for Janell — and making sure the product reflects her. What we learned Maternal mortality is a workflow problem as much as a clinical one. The data needed to save lives is usually already in the chart; what's missing is the right thing surfaced at the right second to the right person. The LLM is the assistant, not the clinician. Deterministic rules for triage, AI for the parts that are actually language work (drafting, summarizing, matching communities). That split made the product more trustworthy and easier to build. Equity has to be designed in, not bolted on. Matching patients to Black-led and culturally-specific support organizations (Black Mamas Matter Alliance, Sista Midwife Productions) is a product decision, not a marketing one. Security review must be continuous. Our first is_provider() check looked correct and was actually a privilege-escalation hole. Running the security scanner on every meaningful migration caught it. A real design system pays for itself fast. Once tokens were in place, every new pathway, card, and flag pill snapped into a consistent clinical look without one-off styling. What's next for GreenCompass Live Epic (FHIR) integration. Replace the mock chart with a real SMART-on-FHIR launch from inside Hyperspace, scoped to the active patient encounter. Bidirectional charting. Write the finalized SOAP note and the triage flag back into Epic as a structured progress note + flowsheet row, with the nurse's edits preserved. Closed-loop escalation. When a red flag fires, page the OB attending and rapid-response team directly from the app with a one-tap acknowledgment, and log time-to-acknowledgment for QI. Postpartum follow-up beyond discharge. Day-3, week-2, and 6-week check-ins via secure SMS for postpartum BP, mood (EPDS), bleeding, and warning signs — the period when most preventable deaths actually happen. Equity dashboard for hospital leadership. De-identified outcome metrics by race, payer, and language to make disparities visible at the unit and system level. Verified provider directory + warm referrals. Move from listing communities to actually scheduling a doula, lactation consultant, or perinatal mental health clinician before the patient leaves the floor. Pilot with a delivery hospital. Move from demo to a measured pilot on a single L&D unit and publish what we learn — including what doesn't work. In Janell's name. Catch what shouldn't be missed.
Built With
- css
- framermotion
- html
- javascript
- lovable
- lucidereact
- markdown
- node.js
- npiregistry
- postgresql
- react18
- reactrouter
- remotion
- stripe
- supabase
- tailwindcssv3
- tanstackquerey
- typescript
- vite5
Log in or sign up for Devpost to join the conversation.