Skip to main content
ECHELON

The architecture standard

Five layers, installed in order, closed into a loop.

Every module we install runs on the same five layers: Connections, Context, Agents, Limits, Console. Work flows down. What the console records flows back up, every week. This page is the standard every install is built to, and the checklist to hold any vendor to, including us.

The closed loop

Five layers. One loop. Nothing you cannot inspect.

Nothing above a layer is trusted until the one beneath is verified. What the console catches feeds back into the limits, the context, and the briefs, every week, as part of running the system.

01Connections

What can it touch?

Read and write into the tools you already run: the CRM, the inbox, the phone line, the calendar, the payment records. Each connection is verified before anything depends on it, and each holds the least access its job needs.

Nothing is ripped out and nothing is migrated. The work lands where your team already looks for it, and every credential is scoped, documented, and revocable in one place.

  • CRM
  • Inbox and calendar
  • Phone and messaging
  • Payments and records
  • Least access

02Context

What does it know?

One record per customer across every tool, plus your price list, your rules, and your policies, kept in your environment. Answers come from your data or they go to a person. Nothing is improvised to fill a silence.

This is the difference between a quoting module that applies your actual price list and a chatbot that invents a discount. The context is yours, it is isolated, and when your prices change the system changes the same hour.

  • One record per customer
  • Price list and rules
  • Policies
  • Grounded answers
  • Escalate on unknown

03Agents

Who does the work?

One agent per job, each with a written brief: what it reads, what it prepares, and what it may do alone. A router in front classifies what arrived and passes it to the agent built for that job. No single prompt guessing at everything.

Every agent starts in review mode. It prepares real work on real events, and a person approves every result until the behavior and the failure paths are understood.

  • One agent per job
  • Written brief
  • Router in front
  • Explicit state
  • Review mode first

04Limits

Where does it stop?

Standing permissions live in one written list, per action and per channel, with the reason for each. Approval gates sit above anything consequential. Narrow authority is granted one action at a time, and never for moving money.

Not on the list means it cannot send. Uncertain matches always go to a person with both records side by side. The right failure is work waiting, not work sent.

  • Permission list
  • Approval gates
  • Narrow authority
  • Escalation rules
  • Never moves money

05Console

Can you see why?

Every step is on the record: what came in, what the agent decided, which tool it called, what came back, and who approved it. The console shows what ran, what it handled alone, what waits on a decision, and who owns it.

Sent means read back and verified. Approved messages send exactly once by design. What the console records feeds back into the limits, the context, and the briefs every week.

  • Audit trail
  • Readback
  • Exactly once
  • Weekly report
  • Named owner

The loop closes

What layer 05 records feeds back into the limits, the context, and the agent briefs every week, as part of running the system.

One run, on the record

Every step, traced.

One routine run, a missed call recovered into a booked appointment, the way it appears on the console. Eight steps, each one inspectable: what arrived, what the agent decided, which tool it called, what came back, and the readback that confirmed it.

Ask any vendor one question: show me the record of why it did that. If they cannot, you are looking at a demo.

Voice

A voice system is a live pipeline, not a chatbot with a phone number.

The phone is where the highest-intent inquiries arrive, and where latency, interruptions, and wrong answers cost the most. Voice runs on the same five layers, built as a streaming pipeline.

  1. 01

    Caller speaks

    Telephony streams the audio in as it happens. No hold music while something processes.

  2. 02

    Heard while speaking

    Streaming transcription produces the words as they are said, so the system is already working before the sentence ends.

  3. 03

    Decide and act

    The router classifies the call, pulls what it needs from context, and acts through the connections: qualifies, answers, books.

  4. 04

    Respond naturally

    Speech streams back with interruption handling. A caller can talk over it and it adjusts.

  5. 05

    Land the outcome

    The appointment lands in your calendar, the record in your CRM with the transcript, and anything below the confidence line goes to a person with context.

In productionA voice intake system built this way answers 200+ calls a day in under a second for a multi-site urgent care group. Read the case study

Isolation by design

Your stack is yours.

Every client runs isolated: separate context, separate credentials, separate configuration. We build inside your accounts, to your security requirements, so the assets are already in your hands.

  • Separate stores per clientContext, transcripts, and configuration are partitioned per company. No shared pools.
  • Least accessEach agent holds scoped credentials to exactly the tools its job needs. Revocable in one place, documented always.
  • No training on your dataYour data runs your system. It does not train anyone's models, ours or anyone else's.
  • Kill switch on every agentAny agent can be paused at once, by us or by you, without taking down the rest of the system.

We sign your NDA before anything sensitive is shared, and a BAA where the work requires it. Sensitive data stays in your environment, high-risk actions wait behind a person, and no model trains on your data. The full readiness list is on the security page.

The stack we build in

TypeScript and Python first. Your stack where it already runs.

The five layers are a standard, not a product. They are built in whatever your codebase already runs on, by a pod sized to the scope, and no new language enters your codebase without a clear reason.

Built with

  • TypeScript
  • Python
  • Node.js
  • Next.js
  • NestJS
  • Remix
  • SvelteKit
  • Vite
  • Turborepo
  • tRPC
  • Prisma
  • Django
  • FastAPI
  • Flask
  • Laravel
  • Symfony
  • Ruby on Rails
  • Elixir
  • Phoenix
  • PyTorch
  • OpenCV
  • Tesseract
  • Celery
  • PostgreSQL
  • pgvector
  • TimescaleDB
  • dbt
  • MySQL
  • SQLite
  • Supabase
  • AWS
  • Google Cloud
  • Azure
  • Cloudflare
  • DigitalOcean
  • Vultr
  • Docker
  • Kubernetes
  • Helm
  • Terraform
  • OpenTofu
  • Ansible
  • Pulumi
  • GitHub Actions
  • GitLab
  • Bitbucket
  • Jenkins
  • CircleCI
  • Argo CD
  • Vercel
  • Netlify
  • Playwright
  • Puppeteer
  • Selenium
Languages
TypeScript and Python first. PHP, Ruby, and Elixir where your codebase already runs on them.
Frameworks
Next.js, NestJS, Remix, and SvelteKit. Django, FastAPI, and Flask. Laravel, Rails, and Phoenix.
AI
LangChain and LangGraph, PyTorch, OpenCV, Tesseract. Retrieval over your own records, extraction, and evaluation.
Data
PostgreSQL first, with pgvector, TimescaleDB, and dbt. MySQL and SQLite where they are already in place.
Cloud
AWS first. Google Cloud, Azure, DigitalOcean, and Vultr, with Cloudflare at the edge.
Delivery
Docker, Kubernetes, Helm, Terraform, Ansible, and Pulumi. GitHub Actions, GitLab, Bitbucket, Jenkins, CircleCI, and Argo CD.
The rule
No new language enters your codebase without a clear reason.

The build standard

Every install ships with all ten.

The checklist a build passes before it goes live in your company, and the checklist to hold any vendor to, including us.

  1. 01Staging before productionNothing ships straight to your customers.
  2. 02Version controlPrompts, routing rules, and configuration are versioned, with rollback.
  3. 03Approval gatesOn every step that moves money or signs your name.
  4. 04Confidence routingAbove the line it acts. Below it, a person decides, with the context attached.
  5. 05Structured loggingEvery step recorded and queryable when you ask why.
  6. 06Alerting wired to usMost issues are fixed before you notice them.
  7. 07Graceful degradationIf a tool is down, the system says so and queues the work.
  8. 08Isolation per clientYour data, credentials, and configuration touch no one else's.
  9. 09Runbook and kill switchDocumented for every agent. Any agent pauses without taking down the rest.
  10. 10Weekly reportWhat ran, what it handled alone, what waits on a decision, and who owns it, plus a live portal.
These five layers run Echelon's own inbound, outreach, and content before they run yours.

Your first department

Bring the process you still run by hand.

One mapping call. You leave with the system design, the price in writing, and the date it goes live. The first build lands in two weeks.

The systems

See the systems we run on ourselves first.

The same infrastructure running Echelon's own inbound, outreach, and content, on sample data.

See the systems