Technology
Standards-first architecture, built for regulated environments.
Java 21 on Spring Boot 3.x, orchestrated by Kubernetes, with an Angular 21 frontend and a model-agnostic LLM gateway. Every channel's semantic contract is an HL7 standard, not a proprietary object model — so customers are never locked in.
Core services
Channel Runtime
Executes deployed channels. Stateless, horizontally scalable, pluggable connector SPI.
AI Agent Service
Hosts the Schema Parser, Mapping Planner, Code Generator, Validator, and Fixer agents on LangChain4j / Spring AI.
LLM Gateway
Model-agnostic façade across Anthropic, OpenAI, and self-hosted Llama/Mistral — prompt caching, rate limiting, cost accounting, provider failover.
RBAC Service
Keycloak-backed SAML/OIDC/LDAP federation issuing JWTs with granular scopes.
Audit Service
Append-only, cryptographically chained event store. Every PHI access and AI-generated artifact is recorded.
Alerting Service
Rule engine with multi-channel notification fan-out and escalation policy evaluation.
Config Service
Git-native configuration store — channels, transformers, and filters committed with PR-based change control.
API Gateway
Spring Cloud Gateway enforcing auth, rate limits, and request tracing.
Data plane: PostgreSQL (metadata, audit), Redis (channel state, cache), Kafka (event bus, high-volume brokering). Observability: OpenTelemetry, ELK, Prometheus/Grafana. Deployment models: single-tenant SaaS, multi-tenant SaaS, on-premise Kubernetes, and hybrid for data sovereignty.
HL7 standards used
HL7 v2.x (2.3–2.9)
Ingestion and egress via MLLP. An internal parser produces a canonical segment tree consumed directly by the Mapping Planner.
HL7 FHIR R4 / R4B
Bidirectional — a destination format for v2-liberated data, and a source for modern EHR APIs (Epic, Cerner/Oracle Health, Meditech). CapabilityStatement-aware.
HL7 CDA / C-CDA
Ingested for care-summary exchange with HIEs. The Schema Parser extracts medication, problem, and allergy sections for FHIR mapping.
Terminology (SNOMED CT, LOINC, ICD-10, RxNorm, UCUM)
Terminology-aware embeddings resolve source codes, flag orphans, and propose canonical codes with provenance.
AI approach
Velantis combines multiple AI paradigms, each applied where it performs best — not a single model used as a hammer for every problem.
Generative AI (LLMs)
Claude, GPT, and self-hosted Llama/Mistral drive the agent pipeline — large frontier models for code generation, smaller distilled models for structured extraction.
Retrieval-augmented generation
Terminology corpora are indexed as embeddings; the Mapping Planner retrieves nearest concepts per field, grounding proposals and cutting code-mapping hallucination from ~18% to under 2%.
Agentic orchestration
A deterministic state machine (Parse → Plan → Generate → Validate → Fix → Approve) — not free-form multi-agent chat — so every transition is auditable.
Program synthesis + execution validation
Generated code runs in a sandbox against sample messages; the Validator’s error trace conditions the Fixer’s next attempt, capped at 5 iterations.
Classical ML for operations
Anomaly detection on throughput, latency, and error rate uses seasonal ARIMA and tree-based classifiers — the right tool for high-frequency time series, not an LLM.
Constrained NLP for operations
Operator queries are parsed by a constrained-decoding LLM into structured queries against metrics and audit stores — deliberately not free-form SQL generation.
Cost & scalability profile
$0.15–$0.80
LLM inference cost per generated channel
~$40k
Replacement labor cost per hand-coded channel — ROI after channel one
1,000+ msg/s
Target throughput per channel instance, runs on a 4-node / 16 vCPU cluster