02 / Technology Five diagrams. One for each procurement question.

Zyron builds AI systems that survive procurement, InfoSec, and the 2am incident.

These are the five diagrams we send to the CTO when they ask how it actually works. Read them in order — each one answers a question the last one raised.

FIG.02 · Enterprise deployment architecture

The map. Zyron as a layer, not a wrapper.

Zyron enterprise deployment architecture Client systems on top, Zyron platform in the middle as a runtime layer, model providers on the bottom, with a governance plane running vertically. FIG.02 / DEPLOYMENT · CLIENT STACK ↔ ZYRON ↔ MODELS LAYER 01 · CLIENT SYSTEMS OF RECORD CRMSalesforce · HubSpotPipedrive · RD Station HRIS & folhaWorkday · BambooHRSenior · TOTVS RM TicketingZendesk · FreshdeskJira · ServiceNow WarehouseSnowflake · BigQueryRedshift · Databricks CommsWhatsApp BusinessSlack · Teams · email ERP · finance · customSAP · Oracle · TOTVS Protheusinternal API · file drops LAYER 02 · ZYRON PLATFORM Zyron runtime layer built + operated by Zyron · deployed in client tenancy or Zyron rack (Houston) Integration gatewayREST · GraphQL · webhooksSFTP · CDC · MCP serversOAuth · SAML · SCIMidempotent · retry-safe Agent runtimeorchestration · tool usememory · retrieval · rerankdetail → FIG.01 Model routingsmall · mid · frontierprovider fallback + budgetper-tenant policypt-BR ↔ en-US routing Governance planeRBAC · tenant isolationPII redaction · policy engineaudit log · LGPD / SOC 2detail → FIG.03 Observability OTelmetricstracesSIEM feed On-call PT + EN Zyron owns LAYER 03 · MODEL PROVIDERS AnthropicClaude · frontier reasoningus-east-1 · sa-east-1 OpenAI · AzureGPT · mid + frontierfallback + audio pipeline Open-weights (self-hosted)Llama · Mistral · Qwenruns on Zyron rack · Houston Embeddings + rerankCohere · Voyage · BGEpt-BR + en-US tuned Governance plane · vertical Zyron does not replace your systems of record. It reads and writes across them, under one governed runtime, with one on-call team responsible. deploy modes: client tenancy (VPC) · shared multi-tenant · Zyron rack (Netrality 1301 Fannin, Houston)
EN

Zyron sits between the systems your company already runs and the models that reason over them. It reads and writes across CRM, HRIS, ticketing, warehouse, and comms through one integration layer, routes each request to the right model tier, and enforces tenant isolation, redaction, and audit through a governance plane that runs the full height of the stack. One vendor is responsible end-to-end.

PT-BR

A Zyron fica entre os sistemas que sua empresa já usa e os modelos que raciocinam sobre eles. Ela lê e escreve em CRM, HRIS, ticketing, data warehouse e canais de comunicação através de uma única camada de integração, roteia cada requisição para o modelo certo e impõe isolamento por tenant, redação de PII e auditoria via um plano de governança que atravessa toda a pilha. Um único fornecedor, responsável de ponta a ponta.

FIG.01 · Agent orchestration pattern

Inside the runtime. One request, one loop, one audit line.

Zyron agent orchestration pattern How a single request moves through classification, retrieval, tool use, model cascade, response with citations, and an audit-log write. FIG.01 / AGENT RUNTIME · ONE REQUEST 01 · INTAKE WhatsApp · Web · API inbound: text / voice / form Auth & tenant scope JWT · tenant_id · locale 02 · CLASSIFY & ROUTE Classifier intent · locale · risk Model router small · mid · frontier pt-BR ↔ en-US aware 03 · REASONING LOOP Agent loop plan → act → observe → repeat Plan step select next tool Memory read episodic · semantic Tool call CRM · DB · calendar · HRIS Retrieval vector · BM25 · rerank LLM call · cascade routed small (class) · mid (Q&A) · frontier (reason) Observation & confidence score ≥ τ → answer score < τ → escalate (see FIG.05) iterate 04 · RESPOND & RECORD Response answer + citations[] locale-matched output Audit log prompt hash · tool trace model · cost · latency citations · confidence retention: 7y (LGPD/SOX) Observability OpenTelemetry · traces + spans LEGEND primary path accent · audit / language iteration τ = confidence threshold, tenant-configurable One request. Up to N tool calls, one to three model tiers, a single citation-bearing answer, and one immutable audit record. median latency (production, Dentflow & Barberflow): 1.4s · p95: 3.9s · escalation rate: 4.1%
EN

Every Zyron agent request runs the same loop: classify, retrieve, call the right model for the job, and record everything. The frontier model is not called on every turn — a small model classifies, a mid-tier answers straightforward questions, and only complex reasoning reaches the frontier. Every step lands in an immutable audit log that survives LGPD and SOX inspection.

PT-BR

Toda requisição a um agente Zyron roda o mesmo ciclo: classifica, recupera, chama o modelo certo para a tarefa, e registra tudo. O modelo de fronteira não é chamado a cada turno — um modelo pequeno classifica, um intermediário responde perguntas diretas, e só o raciocínio complexo chega ao modelo de ponta. Cada passo cai em um log imutável que aguenta auditoria LGPD e SOX.

FIG.04 · Infrastructure stack · Netrality 1301 Fannin, Houston

The rack. We don't rent our runtime from a hyperscaler.

Zyron infrastructure stack Client edge to Zyron API gateway to Zyron agent runtime running on owned rack at Netrality 1301 Fannin Houston, with direct cloud on-ramps to AWS and GCP for regional model providers. FIG.04 / INFRASTRUCTURE · CLIENT EDGE → ZYRON RACK → MODEL PROVIDERS CLIENT EDGE Client network HQ · branches · VPN TLS 1.3 mTLS · mutual auth User surfaces WhatsApp · web · mobile Slack · Teams · voice ZYRON RACK · NETRALITY 1301 FANNIN · HOUSTON, TX API gateway rate limit · auth · quota WAF · DDoS · TLS termination Agent runtime cluster k8s · autoscaling · per-tenant GPU pool for open-weights inference DATA HALL FACTS — 53+ carriers on-net — 27 peering networks — 3 internet exchanges — Direct AWS + GCP on-ramps — N+1 UPS · dual utility feeds — Diesel generators · 65,000 gal fuel — Hurricane-hardened envelope — Outside 500-yr floodplain · Energy Star Storage · vector + object pgvector · S3-compatible · encrypted at rest KMS · secrets · audit sink HSM-backed keys · tenant-scoped Zyron owns the rack. Not a co-tenant on someone else's k8s cluster. Two-country on-call: Houston (US) + Vitória (BR). Pacific + Atlantic coverage. MODEL PROVIDERS (regional) Anthropic · us-east-1frontier reasoningdirect via AWS PrivateLink OpenAI · Azure USmid + frontier · fallbackdirect via Azure ExpressRoute Bedrock · sa-east-1BR-residency frontierfor LGPD-tight workloads Open-weights · on-rack runs in Zyron cluster Llama · Mistral · Qwen Zyron does not rent its runtime from a hyperscaler. The core agent cluster runs on rack Zyron owns, in a Tier III carrier hotel with direct fiber into AWS, GCP, and Azure — one hop to the model, one hop back. Netrality 1301 Fannin · 550 kW critical load · 100% renewable option · SSAE-18 SOC 2 facility · US-CENT-1 carrier hotel Client tenancy option: same runtime, deployed inside client VPC (AWS · GCP · Azure) with private link to Zyron control plane
EN

The core agent runtime runs on rack Zyron owns at Netrality 1301 Fannin, a Tier III carrier hotel in downtown Houston with 53+ carriers on-net, direct fiber to AWS, GCP, and Azure, and enough diesel to hold up through a hurricane. Zyron is not renting compute from a hyperscaler and marking it up — the runtime is Zyron's. If a client requires their own VPC, the same runtime deploys there instead.

PT-BR

O runtime dos agentes roda em rack próprio da Zyron no Netrality 1301 Fannin, um carrier hotel Tier III no centro de Houston, com 53+ operadoras conectadas, fibra direta para AWS, GCP e Azure, e diesel suficiente para aguentar um furacão. A Zyron não aluga compute de hyperscaler e revende — o runtime é da Zyron. Se o cliente exigir VPC próprio, o mesmo runtime é entregue ali dentro.

FIG.03 · Data flow with governance · LGPD / GDPR

Residency is the shape, not a label.

Zyron data flow with residency and governance boundaries Private data flows from source through embedding, indexing, retrieval, prompt assembly, and citation-bearing response. Residency is enforced by a vertical boundary. Provider contracts prohibit training on tenant data. FIG.03 / DATA FLOW · GOVERNANCE · RESIDENCY RESIDENCY: BR REGION (sa-east-1 · São Paulo) RESIDENCY: US REGION (us-east-1 · Houston rack) 01 · SOURCES Private docs · BRHR files · CLT · eSocialpolicies · tickets · calls Structured · BRTOTVS · Senior · ProtheusCDC via Fivetran / native Private docs · USSOP · policies · runbooksConfluence · Drive · SharePoint Structured · USSalesforce · Workday · SAPCDC · warehouse mirror 02 · INGEST · REDACT · CHUNK PII redaction · CPF / RG / CNPJ · tenant-scoped chunking policy engine tags every chunk with tenant_id, residency, classification, retention chunk never leaves BR region · encrypted at rest (AES-256, KMS per tenant) PII redaction · SSN / EIN · tenant-scoped chunking policy engine tags every chunk with tenant_id, residency, classification, retention chunk never leaves US region · encrypted at rest (AES-256, KMS per tenant) 03 · INDEX Vector + BM25 (BR) embeddings pt-BR · pgvector · per-tenant namespace · row-level tenant filter Vector + BM25 (US) embeddings en-US · pgvector · per-tenant namespace · row-level tenant filter 04 · RETRIEVE · RERANK · ASSEMBLE PROMPT Prompt assembler · tenant-scoped · locale-scoped · citation-preserving policy check → retrieval → rerank → context window build → citation index → dispatch 05 · MODEL CALL (zero-retention contract) Provider call · no training on tenant data · region-pinned endpoint · response returns with citations[] → audit log
EN

Private data is chunked, redacted, and indexed inside the same jurisdiction where it was created. A Brazilian dataset never crosses into a US region and vice versa. Every model call runs under a zero-retention contract with the provider, and every response comes back with citations that point to the exact source chunk — so any answer the system gives is auditable to a document a person can open.

PT-BR

Dados privados são fatiados, anonimizados e indexados dentro da mesma jurisdição onde foram criados. Um dataset brasileiro nunca atravessa para região dos EUA e vice-versa. Toda chamada de modelo é feita sob contrato de retenção zero com o provedor, e toda resposta volta com citações que apontam para o trecho exato da fonte — então qualquer coisa que o sistema responder é auditável até um documento que uma pessoa consegue abrir.

FIG.05 · Human-in-the-loop workflow

When the agent isn't sure, it stops.

Zyron human-in-the-loop escalation workflow When agent confidence falls below threshold or a policy trigger fires, the agent pauses and hands a full context bundle to a human reviewer, whose decision is written back to the audit log. FIG.05 / HUMAN-IN-THE-LOOP · ESCALATION · REVIEW · WRITEBACK 01 · CONFIDENCE GATE Confidence & policy gate score ≥ τ → auto-respond score < τ → escalate Also triggers escalation: — high-stakes action (write $) — PII / regulated content — CLT / labor-law answer — unknown tenant policy — user requests human — repeated tool failure 02 · CONTEXT BUNDLE (shipped to reviewer) Reviewer context bundle assembled in ≤ 200ms · deep-linked to source — full conversation transcript — tool calls made · args · returns — retrieved documents (top-N, ranked) — model reasoning trace — confidence score + why it dropped — proposed answer (draft) — policy hits (which rule fired) — similar past cases + prior decisions reviewer decides in seconds, not minutes 03 · REVIEWER DASHBOARD Operator dashboard queue view · SLA timer · assignment Approve draft Edit & send Reject · reply manually Escalate to Zyron reviewer role: client staff (default) · or Zyron on-call (contracted) Zyron on-call answers in PT-BR + EN, from two-country rotation 04 · OUTCOMES (write back to audit + policy) User receives reviewed answer reviewer identity is not shown to end user SLA target: ≤ 90s median · ≤ 5min p95 Audit log records everything who reviewed · what changed · why exportable · retention 7y (LGPD/SOX) Feedback loop edits become eval cases · policy refined threshold τ retuned per topic + tenant A confident agent answers. An unsure agent stops and hands its work to a person who can decide in seconds — because Zyron already assembled the context. production baseline (Dentflow & Barberflow, 2025): 4.1% of turns escalated · 92% resolved by client staff · 8% escalated to Zyron on-call
EN

Confidence is measurable. When the agent scores its own answer below a tenant-configured threshold — or when a policy trigger fires (payments, regulated content, CLT questions, user request for a human) — the loop pauses. A full context bundle ships to the reviewer's queue in under 200ms: transcript, tools used, retrieved docs, model reasoning, proposed draft, and the reason confidence dropped. The reviewer decides in seconds. Every decision writes back to the audit log and refines the threshold going forward.

PT-BR

Confiança é mensurável. Quando o agente pontua sua própria resposta abaixo de um limiar configurado por tenant — ou quando um gatilho de política dispara (pagamentos, conteúdo regulado, dúvidas sobre CLT, pedido do usuário para falar com uma pessoa) — o ciclo pausa. Um pacote completo de contexto vai para a fila do revisor em menos de 200ms: transcrição, ferramentas usadas, documentos recuperados, raciocínio do modelo, rascunho proposto e o motivo da confiança cair. O revisor decide em segundos. Toda decisão volta para o log de auditoria e recalibra o limiar dali para frente.

Ready to talk architecture with our team?

Book 30 minutes