Jason Dicken
Case study 03 · AI agent orchestration / business automationProduction · runs my business daily

Cobblestone CRM

A custom sales CRM where an orchestrator agent, Oreo, directs 20 specialists that research, qualify, draft and schedule. Exactly one module is allowed to contact a customer, and only after a person approves it.

The problem

The work that closes deals (finding prospects, qualifying them, writing proposals, timing the follow-up) was all manual, and it was the first thing to slip when the day got busy.

Agents can do most of that work. But an agent allowed to email a customer unsupervised is a liability, not a feature. The system had to automate the work and keep a person in control of everything a customer sees.

The solution

A CRM built around a seven-stage sales pipeline, with an agent layer on top. Requests arrive from the dashboard chat, from Telegram, or from scheduled jobs. Oreo, the orchestrator, routes each one to the right specialist, and every customer-facing draft lands in an approval queue.

A request comes in dashboard chat · Telegram · 30+ jobs Oreo — the orchestrator holds every tool; the only component with database access A specialist — 1 of 20 no database access; returns the work or a list of what it still needs Draft email, text or proposal — not sent PendingAction — a human approves reviewed in the dashboard or over Telegram; nothing passes on its own One module sends it enforced in code, not by convention
The path a customer-facing message takes, as built and as I operate it. It describes the design; it is not evidence that the implementation is secure or tested.
  • GatedEmail, text messages and proposals, meaning anything a customer receives. These queue as a PendingAction row and wait for a person.
  • Not gatedInternal actions: logging an activity, scoring a lead, scheduling a task, writing a briefing. Nothing in that list reaches a customer.

My role

  • RequirementsEvery workflow, from how leads are found to what counts as qualified, comes from how I sell.
  • ArchitectureOreo-only data access, one sending module, the approval gate, per-agent model tiering. These are my decisions.
  • Agent promptsI wrote and maintain the orchestrator's prompt and the specialists' instructions, and audit them against the code.
  • Development oversightClaude Code wrote most of the implementation to my specifications. I reviewed it and sent back what was wrong.
  • ValidationI test the workflows, read the agent runs, and trace production bugs to their cause.
  • DeploymentI deploy it to its Ubuntu server, operate it daily and fix it when it breaks.

I'm not claiming to have hand-written 361 endpoints. I can specify a system precisely enough that agents build the right thing, catch it when they don't, and run the result as a business depends on it, because one does.

Technology used

  • Python
  • FastAPI
  • SQLAlchemy
  • SQLite
  • APScheduler
  • Anthropic API
  • Claude Opus 5
  • Claude Sonnet 5
  • Claude Haiku 4.5
  • Twilio
  • Gmail API
  • Google Calendar API
  • Deepgram
  • Telegram Bot API
  • JWT auth
  • pytest
  • nginx
  • Ubuntu VPS
  • Android (companion app)
  • Claude Code

Opus orchestrates; Sonnet handles customer-facing writing, money and rules; Haiku handles scoring, extraction and classification.

Key features

  • Sales pipelineSeven stages from prospect to customer, with automated local prospect discovery and lead scoring.
  • Agent orchestrationOreo directs 20 specialists, among them lead qualification, proposals, sales coaching, compliance, marketing and data hygiene.
  • Human approvalEvery email, text and proposal queues for approval in the dashboard or over Telegram.
  • Two ways to call an agentrun_agent_skill for one schema-validated output; ask_agent for a thread where the specialist can ask back for data.
  • Cost controlPer-agent model tiering, cost attribution per run, a daily token budget that stops the fleet, and every run logged.
  • Statement analysisMerchant processing statements read from PDF by a vision model; disagreements with an independent regex check go to a person.
  • Post-call intelligenceTwilio recordings become speaker-separated transcripts, then a structured review written to the lead's timeline.
  • One support queuePOS and loyalty support from email, web forms and in-app chat lands in one place.

Application screenshots

The production CRM holds real customer data, so screenshots will come from a demo dataset. This page will show the real interface only, never mock-ups.

Demo

The repository is private. A live walkthrough is the best way to see the approval queue and the agent runs in the real system.

Development status

  • Working prototype
  • Active development
  • Internal production
  • Customer-deployed

In daily production use, running my own business, and still under active development. No test set exists for the agent fleet, so I make no accuracy claim for it.

Results

The first version was built in about seven weeks. Counted from the running system in November 2026:

AI agents21

Oreo plus 20 registered specialists.

API endpoints361

Across 37 routers.

Database tables64

One schema for sales, support, marketing and operations.

Scheduled jobs33

Registered, a few deliberately switched off.

The honest gap is measurement. The next evaluation project is lead scoring and statement extraction, where being wrong costs money and the right answer can be checked. It will use the same harness as the support assistant.