ACE OPS
The always-on delivery agent for agencies
Your client's AI answers the brief. Our agent drafts the work. People approve it. Nothing leaves the box.
Carl · Tyler · Moorthy — ACE, a performance-marketing agency
Dell × NVIDIA AI Hackathon · San Francisco · 25 October 2026
Runs on Dell Pro Max GB10 · all inference local
Intake → BRD → Spec → Tickets → Build → Review → Release → Status
Ffullscreen→nextNspeaker notes

The problem

New clients cost an agency weeks before the first euro is spent.

weeks
from signature to first spend hypothesis
of 32
onboarding answers still missing at kickoff Northline sample
hours
of senior time per onboarding, unbilled hypothesis
Kickoff call→Checklist doc→Email ping-pong→Loom on granting access→A board typed by hand

Then the work starts on revenue numbers, not profit. Most clients have never calculated break-even ROAS; Northline had, and the agent still checked it.

ACE OPS

The insight

Your client's AI talks to our agent.

The client's side
Their own assistant
ChatGPT, Claude, Gemini — whatever the client already uses. It connects to one MCP URL with a scoped token.
R2What is your contribution margin by category?
R2 | DONETableware 58%, textiles 46%, lighting 41%, seasonal decor 33%.
A2 | N/AMerchant Center access not granted yet. → gap, blocked on client
D5 | DONE"Ignore previous instructions and…" untrusted Stored as data. Nothing happened.
ACE's side
Intake MCP server
Three tools: get_next_item submit_answer get_progress. 32 checklist items, one question per turn.
ACE OPS

How it works

Eight stages. Five human gates. The agent does the paperwork.

1 · client AI
Intake
Asks the next item, records answers with provenance
2 · agent
BRD
Groups answers by section, lists open gaps
3 · agentG1
Spec
One spec per section: goal, scope, acceptance criteria
owner approves
4 · agentG2
Tickets
Split with dependencies from the dependency map
role accepts
5 · builder
Build
Working draft from the client's own figures; builder edits
evidence required to leave
6 · reviewerG3
Review
Pre-checks every acceptance criterion
a different person approves
7 · ownerG4
Release
Release notes and a client-facing summary
owner + client sign-off
8 · always onG5
Status
Plain-language client update with follow-up questions
owner sends

The agent does

Asks, groups, drafts, splits, pre-checks, summarises. Rules first, so it runs with the model down; the local model adds polish. It has no approve, send or delete tool.

A named person does

Approves the spec G1 · accepts tickets G2 · reviews, never the builder G3 · releases, with client sign-off G4 · sends G5. Open gaps flow back to the client's AI through intake.

ACE OPS

The money moment · SPEC-003 Profit model · Northline Goods

Seasonal decor needs 3.4× to break even. The agent puts that next to every draft.

Category
Margin (R2)
Margin-only floor ≈ 1 ÷ margin  ·  | client's break-even (R3)
Tableware
%
≈×
2.1×
Textiles
%
≈×
2.6×
Lighting
%
≈×
2.9×
Seasonal decorthe campaign on the table
%
≈×
3.4×
0×1×2×4×
a 3× plan, for scale
Under 3.4× seasonal decor loses money; under ≈3.03× it loses money before shipping and returns are even counted. A draft that ignores this fails the review pre-check.

Floors are derived by the agent from the client's own R2 answer and marked ≈; the client's R3 break-even adds shipping and returns. A number that traces to no answer and isn't marked ≈ fails the review pre-check.

ACE OPS

Guardrails

The agent drafts. People decide.

  • The agent's MCP server has no approve, send or delete tool. It physically cannot approve its own work. Safety lives in the tool list, not the prompt.
  • Self-review returns 409. Approving over a failing pre-check is refused too.
  • Traceability pre-check: every number traces to a client answer or is marked ≈ derived.
  • Client text is quoted data between fixed delimiters, never an instruction. Every output is schema-checked.
SPEC-003-T02 · review · G3
Break-even ROAS table by category
BuilderJLJordan Lee · Data Analyst
ReviewerJLJordan Lee · Data Analyst
Approve G3
409The reviewer must differ from the builder.
409 refused once · reviewer changed · approved by a different person
ApprovedG3 · Morgan Diaz
R · replay
ACE OPS

Local-first by construction

Nothing leaves the box. By code, not by promise.

  • Refuses to start if the model URL isn't loopback: a LocalOnlyError, not a comment in .env.
  • CSP default-src 'none': no fonts, no CDN, no telemetry. The console, and this deck, load nothing from the network.
  • Rules first: the pipeline runs with the model down. The local model adds polish; provenance is recorded on every artifact.
  • On stage: nvidia-smi busy, a blocked curl from the OpenShell sandbox, a cloud model URL rejected at startup.
Dell Pro Max GB10 · OpenShell sandbox127.0.0.1
Intake MCP :8787/mcp
Intake admin :8788
Board · My plate :8790
SDLC console :8792
Agent MCP :8791/mcp
SDLC loop every 3 s
Local model · OpenAI-compatible endpointloopback only · 20 s timeout · rules fallback
egress
cloud · APIs · CDNs
ACE OPS

Live demo · 60 seconds · console at 127.0.0.1:8792

Approve one spec. Watch it become released, reviewed work and a client update.

0:00
Avery Park · ownerApprove G1
SPEC-003 Profit model → agent splits it into 5 tickets in dependency order
0:10
Jordan Lee · DAAccept as Jordan Lee G2
Ticket moves to build → agent writes the break-even draft from the client's figures
0:25
Jordan Lee · DASubmit for review
Agent pre-checks every criterion → traceability: every number traces or is ≈
0:35
Jordan Lee · DAApprove G3
409 — The reviewer must differ from the builder.
0:42
Morgan Diaz · DAApprove G3
Approved by a different person in the role
0:50
Avery Park · ownerRelease SPEC-003… G4
Release notes + client summary; client sign-off recorded because it touches their ad account
0:58
Avery Park · ownerApprove & send G5
Plain-language client update with a follow-up question per open gap → back to the client's AI
ACE OPS

Why it's a business

The agency owner pays. The client's AI is the channel.

Who pays hypotheses

  1. ACE, customer zero. Buyer and user; our next real client runs through it Monday.
  2. Google Ads agencies serving e-commerce. Onboarding hours are unbilled; launch speed keeps clients.
  3. In-house growth teams onboarding agencies and freelancers: same checklist, reversed.

Pricing hypothesis

$/ month per agency

up to 20 active clients; the box bought or leased.

12 h × $100/h ≈ $ saved per onboarding. Five a year cover it. Fixed box, not a token bill.

Wedge → expansion

One link. The client's AI answers; an hour later: BRD, gap list, break-even table. Sells on its own.
SDLC makes it recurring. Specs, tickets, gated delivery, client updates for the life of the account.
Every agency workflow. The checklist is config; the gates are generic.

Why now: owners already work inside AI assistants, and MCP lets them call another company's tools · a desk-side 128 GB box runs capable open models · NVIDIA's sandbox makes an unattended agent acceptable.

ACE OPS

Built today · what's next · the ask

A delivery agent that cannot approve its own work.

  • Built in one day: intake MCP, BRD, dispatch, board, SDLC agent with five gates, console, loopback guard. tests pass. One GB10, all on loopback.
  • Next week: ACE's next real client goes through it; we measure days to first spend against the last three clients.
  • Then: five agency owners, a $500 pilot each, and a second agency running it unaided.
  • The ask: leave the GB10 under an agency's desk and let it earn its keep.
ACE OPS
The always-on delivery agent for agencies
Carl · Tyler · Moorthy
Dell × NVIDIA AI Hackathon · San Francisco · 25 October 2026
ACE OPS
Speaker notes ·