Amperos Health
Amperos is “healthcare’s first AI-native denial management and revenue recovery platform” (Ashby). Its product is Amanda, a “multi-modal AI coworker” for revenue-cycle management (RCM) that works denied and unpaid insurance claims end-to-end — like a human biller, but across “phone + portals” at once (LLM blog, Boulder blog). The engineering core is exactly that multi-modality: an LLM agent that holds real payer phone calls (voice), navigates payer portals (browser/RPA), reads and writes the practice-management/EMR system (API), then audits its own work before a human ever looks — with offshore RCM billers as the quality backstop.
Vitals: NYC (in-office) · $16M Series A (Bessemer · Uncork · Neo, Apr 2026) + $4.2M seed (Jun 2025) · small team · SOC 2 + HIPAA.
Business context — founders, funding, customers, metrics
- Co-founders: Michal Miernowski (CEO — career in healthcare-focused private equity), Alvin Wu (CPO — product, consumer health → enterprise SaaS), Wilson Wang (CTO — “AI agent developer since early ChatGPT days”) (About). NYC HQ, primarily in-office.
- $16M Series A led by Bessemer Venture Partners, with Uncork Capital and Neo (Apr 22 2026, Series A, Bessemer); builds on a $4.2M seed (June 2025) that launched Amanda, billed as “the world’s first AI biller for healthcare denials and collections” (Series A).
- Customers named publicly: Boulder Care (telehealth SUD), U.S. Urology Partners, EyeCare Services Partners (ESP), Tend Dental, plus an unnamed inpatient provider (Boulder blog, blog index, LLM blog).
- Headline metrics: 22%+ higher claim recovery vs. traditional vendors, 60%+ reduction in AR backlog, 50% lower cost to collect, 500K+ claims processed (home); in the deep-dive, “denials decrease by up to 80% after full deployment,” collectors “working 2–5× more claims within roughly eight weeks,” and Amanda “costs 50–80% less than a human” per action (LLM blog).
- The market framing: U.S. claim denials hit ~12% in 2024, a “$262 billion” problem, amid RCM staffing shortages (Series A).
The heavy lifting
Section titled “The heavy lifting”- One agent, three modalities, one claim. Amanda works a claim across the payer phone line (voice), the payer portal (browser/RPA), and the PM/EMR (API) — “like a human collector … across the PM, payer [portals], and phone” — rather than automating one isolated step and handing back to staff (LLM blog).
- Read-and-reason, not record-and-replay. Unlike RPA that “loops clicks,” the agent “can read on-screen text and understand spoken responses,” so it adapts to portal changes and live call turns — “LLM agents can do the same actions as RPA (click, type), and they can also read and think” (LLM blog).
- Self-audit before handoff. “After every action — portal or call — she pauses to r[eview]”; an “AI auditor” checks each portal action and call summary for completeness, so output is “pre-audited” before a human sees it, and “every call is recorded and transcribed; every portal action is captured and exported to PDF” for proof (LLM blog).
The product is built from job posts, the engineering blog, and the trust footer; the team is a small applied-LLM shop on AWS. Languages/infra beyond Python+AWS aren’t named — reconstructed in Likely internals.
| Layer | Choice | Evidence |
|---|---|---|
| Core language | Python (heavy) for ML / applied-LLM work | AI Research JD |
| Cloud / infra | AWS — dev-ops, observability, monitoring, compliance | Infra JD |
| LLM layer | LLM orchestration, prompt optimization, an evaluation framework, AI infra | AI Research JD |
| Voice | ”state-of-the-art LLMs and speech technology” for payer calls | LLM blog, Full-Stack JD |
| Browser / computer use | browser agents for payer portals | Full-Stack JD, LLM blog |
| Legacy automation | RPA where portals/EMRs have no API | Integrations JD, LLM blog |
| EMR/PM integration | ”APIs, RPA, or agentic flows” into healthcare EMR software | Integrations JD |
| Self-audit | an “AI auditor” reviews each action for completeness | LLM blog |
| Observability | LLM observability + monitoring on the AWS stack | Infra JD |
| Proof artifacts | call recordings + transcripts; portal actions exported to PDF | LLM blog |
| Compliance | SOC 2, HIPAA (monitored by Drata) | home |
| Internal AI tooling | Claude Code, Cursor, Figma Make (named in design role) | Designer JD |
Hard problems
Section titled “Hard problems”The parts an engineer at this company loses sleep over. Public signal is cited (verified); likely approach is labeled speculation — best-practice fill-in, hedged.
| Problem | Why it’s hard | Public signal | Likely approach (speculative) |
|---|---|---|---|
| Real-time voice calls to payers | A non-scripted phone call must hear, reason, and respond fast enough to navigate a payer rep through a denial, reliably, at fan-out | Amanda holds “real conversations — not from a fixed script, but guided by an objective,” with “state-of-the-art LLMs and speech technology” (LLM blog) | Streaming ASR + a fast LLM tier + TTS (OpenAI voice-mode-class), with biller-trained playbooks scoping each call; concurrency scaled per call |
| Brittle, heterogeneous payer portals & EMRs | Hundreds of payer portals and PM/EMR systems change constantly; many have no API; pure RPA breaks | integrations via “APIs, RPA, or agentic flows” (Integrations JD); LLM agents “read and think” vs RPA’s fixed clicks (LLM blog) | Browser agents that read+reason over the live DOM as the adaptive tier, RPA/API where stable; per-system connectors maintained by an integrations team |
| Correctness & auditability in a compliance domain | A wrong claim action loses revenue and is a HIPAA/compliance liability; output must be provable | per-action “AI auditor,” every call “recorded and transcribed,” every portal action “exported to PDF”; SOC 2 + HIPAA (LLM blog, home) | Self-review gate + human-biller QA on low confidence; immutable proof-of-work artifacts per claim; write-back with structured fields |
| Evaluating a non-deterministic agent | Quality of a probabilistic biller across thousands of denial scenarios can’t be unit-tested | AI Research owns an “evaluation framework” and improving the agent “so it acts more human-like and collects more accurate … results” (AI Research JD) | Golden-set + LLM-as-judge over biller-labeled calls/claims; online scoring on recovery rate and override rate before widening autonomy |
Likely internals
Section titled “Likely internals”The infrastructure Amperos doesn’t name publicly, inferred from the stack it does (heavy Python on AWS, multi-modal LLM agent, RCM-biller oversight):
| Component | Likely choice | Basis |
|---|---|---|
| LLM providers | OpenAI + Anthropic frontier models, routed | role wants “o1, OpenAI voice mode, Claude Computer-Use” (AI Research JD); providers not named in production |
| Voice stack | a real-time speech pipeline (OpenAI voice / Deepgram / ElevenLabs-class) + telephony | ”speech technology” for payer calls (LLM blog); vendor unstated |
| Browser-agent infra | a managed cloud-browser layer (Browserbase/Playwright-class) | “browser agents” on payer portals (Full-Stack JD); platform not named |
| RPA tooling | an RPA framework or in-house drivers for no-API systems | ”RPA” named for EMR integration (Integrations JD) |
| Memory / grounding store | a vector index over payer policies + biller playbooks | ”memory & grounding … payer policies, and internal playbooks” (LLM blog) |
| Eval / observability | a prompt-eval + LLM-observability stack (Langfuse/Braintrust-class) | “evaluation framework” + “LLM observability” (AI Research JD, Infra JD) |
| Primary DB | Postgres on AWS | claims/workqueue + relational RCM data; conventional for an AWS shop |
| Auth | enterprise SSO + audited access | SOC 2 / HIPAA, app at app.amperoshealth.com (home); vendor not named |
Architecture
Section titled “Architecture”Amanda’s loop: ground → plan → act (multi-modal) → self-audit → write back
Section titled “Amanda’s loop: ground → plan → act (multi-modal) → self-audit → write back”Amanda is “built on LLMs,” and the loop mirrors an experienced biller. She grounds on context — “prior attempts, payer policies, and internal playbooks” plus the PM/workqueue — then plans against the denial: she “understand[s] denial categories, know[s] which questions” and “thousands of denial scenarios,” and “knows when and how to request reprocessing” (LLM blog). She then acts across modalities — payer portal, phone call, and PM/EMR write-back — and after every action “pauses to review,” with an “AI auditor” checking completeness before handoff. Low-confidence work routes to a human RCM biller; everything is captured as proof (LLM blog).
Mermaid source
flowchart LR classDef io fill:#fdf4e8,stroke:#d97706,stroke-width:1.5px,color:#0f172a; classDef ai fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a; classDef data fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a; classDef human fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a;
Claim(["Denied / unpaid claim<br/>in PM workqueue"]):::io
subgraph Reason["Amanda · LLM agent"] direction TB Ctx("Ground: read PM/EMR +<br/>payer policy + biller playbook<br/>(memory & grounding)"):::ai Plan("Plan: denial category +<br/>evidence required + next action"):::ai Ctx --> Plan end
subgraph Act["Act · multi-modal"] direction TB Portal("Payer portal<br/>browser agent / RPA"):::ai Phone("Payer phone call<br/>voice AI"):::ai Write("PM/EMR write<br/>API"):::ai end
Audit{"AI auditor<br/>action complete & accurate?"}:::data Human("Human RCM biller (India)<br/>AI-quality review"):::human Out[("Write-back to PM/workqueue<br/>+ proof: recording · transcript · PDF")]:::data
Claim --> Reason --> Act --> Audit Audit -->|pass| Out Audit -->|low confidence| Human --> Out Human -. "labels feed back" .-> Plan Out -. "not resolved -> next touch" .-> ClaimIntegration surfaces: where Amanda plugs in
Section titled “Integration surfaces: where Amanda plugs in”The product’s reach is defined by what it can touch. Integrations are built “with healthcare EMR software using APIs, RPA, or agentic flows” (Integrations JD); Amanda “orchestrates actions across systems and writes back structured updates” to the PM/workqueue, and the captured data drives analytics — customers “gain clear visibility into payer- and code-level patterns” (LLM blog). Deployment is configured per customer: “your Amperos deployment team maps your exact workflows, payor priorities, and standard operating procedures into the system” (Solution).
Mermaid source
flowchart LR classDef sys fill:#eef2f8,stroke:#94a3b8,stroke-width:1.5px,color:#0f172a; classDef ai fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a; classDef data fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a; classDef human fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a;
subgraph Systems["Systems Amanda works across"] direction TB PM("PM / EMR<br/>API · RPA · agentic"):::sys Portal("Payer portals<br/>browser agents"):::sys Phone("Payer phone lines<br/>voice AI"):::sys end
Amanda("Amanda<br/>multi-modal AI coworker<br/>orchestrates across systems"):::ai
subgraph Plane["Amperos platform · AWS · SOC 2 · HIPAA"] direction TB WB[("Structured write-back<br/>to PM / workqueue")]:::data Proof[("Proof artifacts<br/>recordings · transcripts · PDFs")]:::data Analytics("Analytics<br/>payer- & code-level patterns"):::data end
Oversight("Human oversight<br/>RCM billers · AI-quality (India)"):::human
Systems <--> Amanda Amanda --> WB Amanda --> Proof WB --> Analytics Amanda -. "low-confidence escalations" .-> Oversight Oversight -. "corrections / training signal" .-> AmandaTeam & process
Section titled “Team & process”A three-founder team — CEO Michal Miernowski, CPO Alvin Wu, CTO Wilson Wang (“AI agent developer since early ChatGPT days”) — building in-office in NYC (About).
| Role | Person | Source |
|---|---|---|
| Co-founder / CEO | Michal Miernowski — healthcare private equity | About |
| Co-founder / CPO | Alvin Wu — consumer health → enterprise SaaS | About |
| Co-founder / CTO | Wilson Wang — AI agent developer since early ChatGPT | About |
The engineering org is a compact applied-LLM shop ($170–300K + equity): AI Research (LLM orchestration, eval, AI infra — “heavy Python”), Full-Stack (voice AI, browser agents, generative features), Integrations (EMR via API/RPA/agentic), and Infra (AWS dev-ops + LLM observability) (AI Research JD, Full-Stack JD, Integrations JD, Infra JD). There is no in-house research org and no QA team — quality is the “AI auditor” plus offshore RCM billers (LLM blog, Ashby). Delivery is design-partner-led: an “AI Deployment Strategist” and a deployment team that “maps your exact workflows, payor priorities, and standard operating procedures into the system” (Solution, Ashby) — configure-per-customer, then let Amanda work, with humans auditing as autonomy widens.
Sources
Section titled “Sources”Reconstructed from public sources only — no insider information. Crawled 2026-06-09 via Chrome MCP (logged-out) + the Ashby posting API. First-party (amperos.com, the Amperos blog, Amperos’s Ashby board) prioritized; Bessemer post labeled third-party. Claim tiers: verified (stated on a public page, linked) · inferred (reasoned from a cited signal, confidence flagged) · speculative (best-practice fill-in, labeled). Links are live; pages change, so the supporting quote for each claim is kept in this repo’s evidence map (evidence/amperos-health-evidence-map.md).