Multi-agent RevOps review · Applied AI

RevOps Software Factory

Review a synthetic revenue batch with a multi-agent backend and live cost tracking.

Buyer & pain

Applied AI Engineer / AI Solutions Architect / RevOps Engineer — revenue data rots silently: duplicate accounts inflate the pipeline, attribution credits the wrong contact, closed-won deals carry no close date, enrichment goes stale. A human analyst auditing a batch is slow and expensive; nobody audits every sync. The review is the constraint.

Problem

One generalist LLM prompt over a whole CRM export is expensive, unfocused, and unverifiable. There's no cost-bounded, specialized review layer that catches the specific failure modes of revenue data and returns one trustworthy verdict.

Before

A 29-record quarter-end pull goes into the forecast unaudited. Two duplicate Vertex accounts double-count a $410k deal; a closed-won deal has no close date; attribution credits a contact at the wrong company. The number is wrong and nobody knows.

After

● REQUESTED CHANGES · 4 agents · $0.0x — a coordinator dispatches dedup / attribution / stage-logic / enrichment specialists, fuses their findings, drops the noise, and returns one approval-biased verdict with a live cost ledger. The blast radius sets the team size.

What the demo proves

  1. Tokenomics — cheap specialist models, a mid-tier coordinator, a live ledger pricing the run against an analyst-hour. Knowing not to put the top model on everything is the signal.
  2. Many specialists beat one generalist — each agent reads only its record slice (diff-partitioning) and emits structured-severity findings; the coordinator dedups and reasonableness-filters into one verdict.
  3. Risk-scaled compute — trivial batches get 2 agents and a cheaper coordinator; full batches get the dream team. Don't send the dream team to review a nightly sync.
  4. Built for failure — per-call timeouts, single failback to a previous-gen model, a KV budget + per-IP rate limit, and failback to the last cached run. Production-grade, not a one-shot.

How it works

  1. Pick a synthetic batch — Trivial / Lite / Full HubSpot-shaped records with planted, gradeable issues. No real CRM data.
  2. Real model calls, server-side — specialists (Haiku) + coordinator (Sonnet) via a Cloudflare Pages Function; the client only streams the JSONL trace.
  3. Cache + budget guard — most clicks serve a cached real run (instant, ~free); the UI labels every run fresh / cached / failback so the "live" claim stays honest.
  4. Public repo — github.com/dallascrilley/revops-factory-demo: the engine, agents, and fixtures, unit-tested.

Integration surface

Astro 5 front end + a Cloudflare Pages Function backend. Deterministic core (risk classifier, fusion, cost, cache-key) is pure and unit-tested; LLM calls are isolated behind a transport with a fixture mode. Deploys as a static Pages subpath with a KV namespace.

Skills proven

Roles this maps to

Try the live demo → dallascrilley.com