BUILT · PRE-PRODUCTION

Atlas PM

An operations platform for commercial moving & logistics, with an embedded AI operator.

Atlas PM was designed from months of field research inside a real commercial moving and logistics operation. We built the full platform — field walkthroughs, quoting, client e-signature, job execution, automatic invoicing, warehouse and payroll — and validated it end-to-end against demonstration data. It has never run a live operation. That is the next test this research points to: hardening the confirm step described below from a prompt-level instruction into an independent backend state machine, before anything real touches it.

Scale
~62,000 lines of code (Python + TypeScript), 29 migrations
Stack
FastAPI + SQLAlchemy (async) + PostgreSQL · Next.js + React + Tailwind
Pipeline
Walkthrough → Quote → e-Signature → Job → Auto-invoice → Warehouse/Payroll
Field
Mobile PWA, route-optimized crew dispatch
Tenancy
Multi-tenant from the base — tenant_id on every model
Status
Complete system, validated end-to-end with demonstration data — pre-production
How it works — under the hood

Read across every module

Atlas answers questions over jobs, quotes, invoices, warehouse and payroll — with tenant isolation enforced at the tool layer, so the AI cannot cross data boundaries by construction.

Proactive scanner, coded rules

Business rules run against the live schema — overdue invoices, jobs pending confirmation, unapproved timecards — each with its own threshold and severity, surfaced without being asked.

Bounded tool-calling loop

A hard cap on iterations, explicit handling for rate limits and permission errors, required-field and format validation before anything reaches the database.

Two-step-confirm writes

Every write action returns a preview first and executes only after explicit human confirmation. That is the lab’s thesis — AI as a partner, not a slave — implemented in a full operations platform, not a toy example.

The full lab note — designing an assistant that acts, but only with permission, and the honest limit of confirm-by-prompt — is published in the research section.

→ Lab note: Embedded AI in a real operation