Evidence before eloquence
Claims about a prospect must map back to captured public evidence. Unsupported specifics are rejected or quarantined instead of being polished into confident copy.
An AI revenue system that can fail safely.
A continuously running, owner-operated agent stack for discovery, evidence-grounded drafting, outreach operations, payment verification, fulfillment, and recovery. The useful part is not the prompt—it is the control system around it.
The operating loop
DiscoverFind public, relevant purchase intent and preserve the source evidence.
EvaluateRequire structured output, source support, policy checks, and deduplication.
OperateQueue outreach, classify replies, monitor delivery, and route exceptions.
Close & fulfillCount only confirmed payment events, then create a traceable delivery job.
Where it earns trust
State before retries
SQLite-backed state, idempotency keys, and disposition history prevent duplicate work—including the quiet failure where a rejected target keeps reappearing.
Evaluation before delivery
JSON schema validation, deterministic gates, a separate review pass, and human sign-off for risky external actions catch malformed or wrong output before a client sees it.
Truth before dashboards
Revenue is recorded only from confirmed Stripe payment events. Sends, bounces, replies, payments, and fulfillment are separate states rather than one optimistic funnel number.
Failure modes designed for
Model timeouts or invalid JSON
Bounded retries, schema checks, safe fallbacks, and explicit failure records keep one weak generation from corrupting the queue.
Hallucinated prospect observations
Verbatim evidence membership checks and an independent reviewer force grounded output; uncertain material does not graduate to send-ready.
Authentication and delivery faults
OAuth health checks, circuit breakers, bounce suppression, and watchdog restarts turn silent infrastructure failure into visible operator action.
Non-technical operator handoff
Human-readable status, bounded queues, saved artifacts, and clear escalation states let an owner understand what ran, what failed, and what needs approval.
Production stack
Custom PythonSQLiteLocal Qwen modelsGPT/Codex reviewGmail OAuthStripe eventsCloudflare Pages + WorkermacOS launchdWatchdogs + KPI snapshots
What this proves—and what it does not
This is a live operating system on the owner's Mac, not a slide-deck architecture. It has processed real outreach, detected bounces, maintained state across recurring runs, and reconciled payment truth. It is not presented as proof of revenue success: acquisition performance and system reliability are measured separately, and the current commercial experiment remains in progress.
How I would adapt the pattern to commerce
Start with a deterministic onboarding state machine: least-privilege credentials, shop validation, brand constraints, a dry-run listing audit, evaluator thresholds, human approval, execution, and a complete audit trail. Add generative capability only after state, idempotency, observability, and rollback behavior are explicit.