Yumi Health

Inspiration

What it does

Here's a draft tailored to what we actually built. Edit to taste:

Inspiration I'm an NJ-licensed insurance agent. Watching consumers try to shop for Medicare and ACA plans, I saw the same pattern over and over: people don't fail to compare plans because the choices are too hard — they fail because the language is impenetrable. On-exchange, off-exchange, APTC, SEP, 48-hour SOA cooldown. Kaiser Family Foundation reports that 29% of insured adults say their plan doesn't meet their needs, and $31.9B per year is lost to suboptimal selection. 73% of Medicare beneficiaries never compare plans during open enrollment. We wanted to fix the layer that breaks first: explanation.

NJ specifically is a State-Based Marketplace, which means subsidized ACA enrollment legally has to route through a licensed advisor — a fact that confuses 95% of shoppers who expect to click an Enroll button. Existing tools either pretend that constraint doesn't exist or pretend the consumer should understand it. We decided to build the third option: tell the truth, in plain English, with the advisor relationship baked in as a feature, not a friction.

What it does Yumi Health is an NJ-focused Medicare + ACA enrollment platform with three connected surfaces:

Consumer plan finders at /medicare-finder and /aca-finder — Sunfire-style detail, real carrier wordmarks, drug pricing matrix (5 tiers × 30/60/90-day supply × preferred/standard pharmacy), per-plan in-network provider check, chronic-condition tagging, opt-in Yumi cost prediction sidebar with probability bands. A plain-English ACA eligibility screener at /aca-eligibility — 7 questions, 4 plain-language verdict screens that translate marketplace regulation into "you qualify for ~$280/mo in financial help; NJ requires this go through an advisor — we handle it for free, the carrier pays our fee." A licensed-agent CRM — leads, applications, SOAs, paste-helpers for the manual Get Covered NJ submission workflow, audit logs, HealthSherpa webhook attribution, automated comms with product-aware confirmation emails. Backed by CMS Public Use Files + CMS Plan Benefit Package 2026 data for 7,420 plans nationally, HealthSherpa One API for live ACA quotes, NPPES + RxNorm for federal provider/drug lookups, and FHIR Provider Directories from 7+ carriers per the CMS Interoperability Rule.

How we built it Next.js 16 + TypeScript + Tailwind on the front, Azure Container Apps + Azure PostgreSQL + Prisma v7 (with @prisma/adapter-pg) on the back. SignalR for realtime, Azure Blob for compliance docs, Azure Communication Services for email. NextAuth.js v5 for agent auth with Microsoft Entra ID step-up for QuickBooks payouts.

Live data: HealthSherpa One API for ACA quotes (CMS-certified DE Entity); custom ingest scripts for CMS PUFs and CMS PBP TXT files (the parser does per-section column mapping for B1a inpatient hospital, B4 ER/urgent, B7 PCP/specialist, B10 ambulance, B13 supplemental, B16/17/18 dental/vision/hearing); NPPES and RxNorm typeaheads; per-carrier FHIR Provider Directory cache with 30-day refresh; Brandfetch API to fetch and self-host real carrier wordmarks. Monthly auto-refresh cron via Azure Container Apps Jobs hitting /api/cron/refresh-cms-puf so plan data never goes stale by more than 35 days.

The cost simulator engine is a novel Monte Carlo system with a CSEFL feedback loop (Patent filing in progress — see docs/cost-simulator-methodology.md) that integrates a Gaussian copula for correlated comorbidities, a Markov chain for monthly health state transitions, Cox proportional hazards, Bayesian posterior updating, and behavioral adherence modeling — producing per-plan probability distributions instead of single point estimates.

Challenges we ran into NJ is a State-Based Marketplace. Subsidized ACA enrollment cannot be self-submitted regardless of which API we use; it must route through Get Covered NJ via a licensed advisor. We embraced the constraint by building a paste-helper for the manual workflow and explaining the rule to consumers up front instead of hiding it. Jargon as the real product problem. "On-exchange / off-exchange" means nothing to 95% of shoppers. We swept the entire codebase to translate it into "with federal financial help / private plan," and the screener verdict screens explain in three sentences what regulation makes a particular path mandatory. CMS data sources are real but inconsistent. Plan Benefit Package files ship as Microsoft Access exports with per-carrier column quirks; cost-share is encoded as min..max ranges where consumer display wants max; CMS quarterly URLs rotate; some plans use Section B13 catch-all slots for completely different benefits identified only by free-text description. We solved each in the ingest layer instead of papering over with fake data. Carrier APIs are gated. Even with RTS status for Humana / UHC / Aetna / Cigna, individual carrier APIs aren't reliably available. We pivoted to FHIR Provider Directories (federally mandated) + per-carrier "Find a Doctor" deep links as a guaranteed fallback so every plan card has actionable network info. Container Apps internal routing. The cron job kept getting 503s calling its sibling app via the public FQDN — turns out service-to-service within an env has to use the internal FQDN. Compliance everywhere. Every estimate is labeled as an estimate; the Monte Carlo simulator is opt-in with prominent "predicted estimate, not insurance guarantee" disclaimers; no fake data anywhere; monthly refresh badges visible to consumers. Accomplishments that we're proud of Real CMS PBP data loaded for 7,420 plans, verified field-for-field against Sunfire's display for UHC NJ-7: PCP $0, Specialist $55, ER $130, Urgent $50, Ambulance $275, Inpatient $550/day × 5 days, Dental $2,500, Vision $300. Every number is a real CMS-reported value, not a synthetic placeholder. A patented Monte Carlo cost simulator that captures stochastic dominance, tail risk, and behavioral feedback — the first consumer-facing health insurance comparison tool to model uncertainty instead of single-point estimates. A 7-question screener that translates 45 CFR §155.305 marketplace eligibility into language a 67-year-old can read, with explicit explanations of why an advisor handles their enrollment. 17 real carrier wordmarks self-hosted via the Brandfetch authenticated API — zero third-party client-side dependencies, no data leak to logo CDNs. End-to-end HealthSherpa webhook integration verified with correct NPN attribution, lead creation, status flip to enrolled, and audit trail. Monthly automated CMS data refresh scheduled via Azure Container Apps Jobs — consumers see a "CMS data refreshed (Nd ago)" badge in the footer, so the platform holds itself publicly accountable to its freshness commitment. What we learned The hardest part of healthcare comparison isn't data — it's translation. Build the explanation layer first, the API layer second. CMS data is genuinely free but takes real infrastructure investment to operationalize. The companies you'd pay $50k/yr to license this from did the same work. For a small FMO competing against Pinnacle, the moat is CRM polish + agent workflow, not data access. Both will eventually be commoditized. Direct carrier API access is gated and unreliable. FHIR Provider Directories (federally mandated under CMS Interoperability Rule) + carrier consumer-directory deep links are the practical fallback. Stochastic modeling for consumer cost prediction works — but it has to be wrapped in legal disclaimers and opt-in toggles, not pushed past consumers as a guarantee. What's next for Yumi Health HealthSherpa off-exchange enrollment API integration — ships ~week of June 12, 2026. The wizard is already structured for the schema flip. Per-plan provider search — NPPES typeahead joined with our per-(NPI, payer) FHIR cache so consumers can browse "who's in network for this plan" instead of just verifying their existing doctors. More FHIR endpoints seeded — Clover, Braven, Longevity, AmeriHealth, Wellpoint to extend in-network verification to every NJ plan. Spanish and Chinese translations — NJ is one of the most linguistically diverse states; the screener and plan finder both need to ship in at least Spanish for OEP 2026. National expansion as an FMO — competing with Pinnacle, recruiting NJ-licensed agents into a Yumi-branded CRM with the cost simulator and paste-helpers built in. The patent application supports the moat. Yumi Forever integration — a sister project providing senior services (in-home care, transportation, etc.) that ties into the Medicare enrollment funnel.

Built With

• a
• and
• carlo
• carrier-fhir-provider-directories
• cms-public-use-files-+-plan-benefit-package-2026
• cost-simulation
• custom
• monte
• next.js-16-+-typescript-+-tailwind-on-azure-container-apps-+-postgresql-(prisma-v7)
• nppes
• powered-by-healthsherpa-one-api
• rxnorm