Carelyze Platform
Designed and built an AI agent that autonomously executes healthcare SOPs — from the ReAct loop and context management to the landing page and email templates.
The Problem
Small independent clinics struggle to close care gaps and schedule preventive visits at scale. Manual outreach is slow, error-prone, and can't keep up with compliance requirements — HIPAA for patient data, TCPA for communications. Carelyze automates the entire patient engagement lifecycle, from intelligent outreach campaigns to AI-powered clinical operations, so clinics can focus on care, not admin.
What I Built
Carelyze is an AI-powered SaaS platform for healthcare operations. As cofounder, I worked across every surface: designed and built the AI agent architecture, contributed to the FastAPI backend (45+ API routes, 58 services), contributed to the Next.js frontend (91+ components), built the marketing website solo, designed all email templates, and learned DevOps to deploy everything on Kubernetes.
Dashboard & Case Management
Visual Workflow Builder
AI Agent with ReAct Loop
Autonomous agent that classifies tasks, generates plans from SOP templates, executes tools with policy-aware risk management, and maintains context via embeddings-based memory. Human-in-the-loop approval gates for high-risk actions.
Campaign Orchestration Engine
Graph-based state machine driving multi-step patient outreach (SMS + email) with condition branching, wait nodes, consent capture, and appointment scheduling via nanosites. TCPA/HIPAA compliance handled automatically.
Workflow Builder & SOP System
Visual ReactFlow-based designer where users define SOPs as descriptive blueprints. 20+ node types. The agent interprets these SOPs and adapts execution to context — a 5-step SOP can handle edge cases the author never anticipated.
Marketing Website (Solo Build)
Email Templates (Solo Design)
Design Decisions
Workflows-as-SOPs, not executable automations
Instead of rigid executable workflows, we made workflows descriptive process blueprints. The agent interprets and adapts to context rather than blindly following steps. Core philosophy: trust the LLM to be intelligent — our job is to provide the right context, tools, and guidelines.
Policy-first tool execution with risk levels
Every tool has a risk level (LOW / MEDIUM / HIGH). The policy engine evaluates each invocation against org-specific thresholds before execution. You can't let an agent send SMS to 10,000 patients without a human gate, but you also can't make analysts approve every file read.
One ticket, many agent runs
A single WorkItem can have many AgentTask executions. For email chains, instead of creating a new ticket per email, all agent runs link to the same ticket. The human sees one coherent thread with full execution trace — dramatically reducing noise for healthcare analysts.
How I Built It
ReAct Agent Loop with Context Builder
Multi-phase agent (Clarify → Plan → Execute → Reflect → Report) in Python/FastAPI. ContextBuilder assembles phase-appropriate context from conversation history, SOP guidance, tool manifests, and pgvector memory. PolicyAwareReasoner scores decisions — heuristic shortcuts for deterministic steps, LLM reasoning for ambiguous ones.
Multi-Tenant Campaign State Machine
CampaignExecutionEngine processes enrollment state transitions every 30s via APScheduler. Graph-based campaigns with 7 typed nodes and conditional edges. Optimistic locking prevents concurrent processing. Organization-scoped via ContextVar-based tenant isolation with Row-Level Security.
Kubernetes Deployment on K3s
Full stack on K3s with NGINX Ingress, Let's Encrypt TLS, and automated ECR image update cron jobs. Six services across 4 domains. PostgreSQL with pgvector, RabbitMQ, and Redis with persistent volumes and resource limits.
AI Agent ReAct Loop
Task → Observe → Classify → Plan → Execute (with policy gates) → Reflect → Report → Save to memory
Outcome & Impact
Full-stack platform shipped to production — End-to-end from agent architecture to Kubernetes deployment — 45+ API routes, 58 services, 91+ frontend components, 45+ database models, all running in production on K3s.
AP Pilot ready for clinic onboarding — Two-campaign architecture (acquisition + reminders) with dual entry paths, compliance guardrails (TCPA/HIPAA), and automated appointment scheduling via nanosites.
Autonomous agent executing healthcare SOPs — Agent classifies tasks, generates plans, executes tools with policy enforcement, and maintains long-term memory — with full audit trail for healthcare compliance.
Wore every hat — Designed the agent architecture, built frontend views, contributed to backend services, created the marketing site and all email templates, and deployed infrastructure — cofounder doing whatever it takes to ship.