Back to work
Artificial IntelligenceHealthTech Security
Lumi

Intelligent care with privacy-first AI.

Architecting a secure, AI-powered clinic management platform for therapists, utilizing hybrid LLMs and a military-grade privacy pipeline.

In this study
01 · Baseclinic management
02 · Plinthfor therapists
03 · Stemhybrid llms
04 · Collarprivacy pipeline
05 · Mitrepii scrubbing
06 · Finialprivacy-first ai
Hybrid Intelligence

Hybrid Intelligence Engine

Lumi isn't just a wrapper. We engineered a dual-model system. Gemini handles broad reasoning and complex summarization, while a lightweight Custom LLM, fine-tuned on anonymized clinical datasets, handles domain-specific terminology and structured data extraction.

We engineered
  • Context-Aware Routing:An intelligent router decides which model to query based on request complexity and cost.
  • Next.js + Node.js:A high-performance, type-safe full-stack environment using Prisma ORM for seamless data compatibility.
01 · Basenext.js + node.js
02 · Chestprisma orm
03 · Throatcontext-aware routing
04 · Jawpii redaction layer
05 · Crestgemini pro
06 · Earcustom llm
Therapist Session NotesRaw Audio/Text
PII Redaction LayerNER / Regex
Gemini ProComplex Reasoning & Insight Generation
Custom LLMFine-tuned on Clinical Methodology
HIPAA Compliance

Zero-Trust Privacy Pipeline

Patient trust is paramount. We implemented a rigorous anonymization pipeline that ensures no Personally Identifiable Information (PII) ever reaches the LLM providers.

Raw Input

Session Transcript

PII Scrubbing

NER + Pattern Match

LLM Processing

Anonymized Context

● Data at Rest: AES-256 Encryption   ● Data in Transit: TLS 1.3

Cloud Infrastructure

High Availability & Cost Efficiency

Deployed on AWS using Terraform for reproducible infrastructure. The architecture balances reliability with startup-friendly costs.

Multi-AZ ECS Fargate

Compute workload distributed across 3 Availability Zones for fault tolerance. Utilizing Fargate Spot integration for 70% cost reduction on non-critical tasks.

Aurora Serverless V2

PostgreSQL database that instantly scales compute capacity to match demand, dropping to minimum ACUs during off-hours to save costs.

IaC & Pipeline

Terraform PlanInfrastructure changes visualized
CI/CD (GitHub Actions)Lint → Test → Build Container
Terraform ApplyAtomic state update & Rollout
Product Engineering

Built for Focus & Scale

We designed Lumi to disappear. The interface is intentionally minimalist to reduce cognitive load for therapists during sessions, while the backend is architected for massive concurrent throughput.

Clinical “Zen” UI

  • ›Dark Mode First: Reduces eye strain during late-night charting sessions.
  • ›Focus States: Input fields enlarge and dim background noise when active.
  • ›WCAG 2.1 AA: High contrast ratios for accessibility compliance.

Frontend Engineering

  • ›React 18 + Vite: Instant HMR and concurrent rendering features.
  • ›Tailwind CSS: Atomic utility classes for a design system that scales.
  • ›Real-time State: Optimistic UI updates for instant interaction feedback.

Backend Architecture

  • ›Node.js Microservices: Decoupled services for auth, transcription, and LLM via gRPC.
  • ›Event-Driven: RabbitMQ for asynchronous processing of long-running AI tasks.
  • ›Prisma ORM: Type-safe database access with automated migrations.

Have a system that needs to hold up?

Start with a free cost and governance review. You learn where you're overpaying and where you're exposed, before you commit to anything.