Most marketing leaders bought AI.I built mine,and it runs the company.
I'm Josh Cook. In 19 months I took Buzzword Strategies from zero to
$1M+ ARR with no outside capital, by building the multi-agent platform that
does the delivery. 24 named agents, 169 packaged skills, a local GPU inference
tier and a client-facing conversation AI, all in production for 21 accounts on a team of eight.
Every figure on this page came off a system you can watch running.
Every role in this agency has a name. Most of them are software.
Buzzword delivers paid media, creative production, CRM automation, lifecycle email, landing pages,
lead magnets and performance reporting for 21 clients. That workload normally needs thirty people.
It runs on eight, because the delivery layer is software, and the software is organized the way a
good agency is: specialists with clear remits, a router above them, and a gate before anything ships.
The EA
Text-message interface to the whole system
An agent I run the company through from my phone. I text it, it works out what I meant, and it
acts. There is no menu to learn and no command syntax to remember.
The conversational turn is read-only
by design, so asking it a question can never change anything as a side effect. Work that changes state
leaves as a dispatch into its own lane with its own cap, log and timeout. Anything a human will read
leaves as a draft and comes back for approval.
A guardian process keeps the lane alive with an outbound heartbeat. An inbound webhook
cannot cold-start a machine that has already stopped listening.
Amanda V6 & Jess
Conversation AI, client-facing
Amanda is the conversation AI deployed into any client CRM instance by
slug: she books, qualifies and follows up inside their pipeline, tuned per account.
Jess is the setter. She runs on one canonical persona file so she can never
contradict herself across channels, and she carries an absolute honesty rule: asked sincerely whether
she is a person, she says she is an AI assistant, every time.
The persona is there to take the sales pressure out of a first message. If it ever left
someone believing they had spoken to a person, it would have failed.
BuzzRouter
The orchestrator. Reads any incoming task, routes it to the right specialist, and owns the hand-offs between them.
OpsChief of staff
Runs the morning sweep, drafts the daily brief, maintains per-client state and keeps the other agents honest about what happened overnight.
GateApprovals
The human checkpoint. Owns an append-only approval queue that any specialist can write to and nothing ships past.
MayaCreative director
Writes every word a human outside the company reads. Ad copy, landing pages, nurture email, conversation scripts.
QuinnAuditor
A binary gate between creative and deployment. Audits every outbound asset against compliance and brand rules, then returns pass or fail.
SageAsset design
Turns a creative brief into finished production assets. Most of the work is in the generation prompts, which decide whether an image is usable or just decorative.
JillianMedia buyer
Runs the Meta campaigns: creative production, ad copy, state-by-state targeting, budget pacing and top-of-funnel testing.
PaigeLanding pages
Builds and deploys landing pages and resource hubs, from a single conversion page to a full Next.js hub.
KrisAccounts
Drafts client email in my voice, builds the narrative for performance reports, and carries onboarding follow-up.
ScoutQA
Drives a real browser to test the whole funnel end to end: ad to landing page to CRM, verifying the conversation AI actually fires.
LexCRM fleet
A router over four CRM specialists: one builds pipelines and workflows, one audits live state read-only, one deploys the conversation AI, one runs day-to-day ops. Fires into any of 23 client instances by slug.
The FactoryResearch & publishing
Six agents that produce a 28-to-40 page evidence-backed buyer's guide: harvester, researcher, scribe, compositor, adversarial critic and citation librarian. Every claim traces to a fetched, verified source.
02 · The proof
Anyone can describe an AI strategy. This is mine, running, with a cost ledger.
This is the live operations console, captured from the machine the platform was built on. Over seven
days it routed 2.17 million tokens to local GPU inference instead of a metered cloud
API, with model routing per task type, automatic context sizing and a savings ledger.
Bulk extraction, classification and summarization run on hardware I own, and the metered models are
kept for the judgment calls. That gap is most of what separates using AI from engineering with it.
Live console, captured today. Left: resident models on a 16GB RTX 5070 Ti. Right: seven days of local
inference broken out by lane, with the routing map beneath. Client identifiers and account spend redacted.
1,654Automation scripts
Wired into delivery, from creative QA to budget pacing.
169Packaged skills
Composable capabilities the agents call.
30MCP integrations
CRM, ad platforms, analytics, billing, storage.
23Isolated instances
One CRM per client, each with its own creative memory.
8Local models
Routed by task: reasoning, extraction, vision, OCR, embeddings.
3Quality gates
Compliance, brand and AI-detection, enforced in the pipeline.
The QC gate
Why the volume is safe to ship
Every static ad passes a locked final gate before it exists as a deliverable. A deterministic lint
checks brand hex values, banned phrasing, duplicate copy, logo placement and client-data leakage. Anything
that fails is automatically recreated and re-checked, looping until clean.
Quality at this volume cannot be a review step. It has to be a property of the pipeline.
The humanizer
Nothing ships reading like a machine
A detection layer sits on every publish surface: ads, email, landing pages, ebooks, social. It flags
AI writing tells and blocks the asset. A warning would just get waved through.
It was tuned the hard way. A false positive that stops a real campaign is worse than no rule, so every
hard rule is blast-radius tested against live batches before it ships.
03 · The products
Named systems, with real users and real consequences.
Almost everything below exists because something went wrong in front of a real customer, and the
fix was a rail that makes that failure impossible to repeat. The rail became the product. Two of
them are the same architecture arrived at twice: a model that forms the judgment, wired to
code that holds all authority to act.
The screenshots are the running systems, captured 2026-08-18. Prospect names, client names, account
identifiers and balances are redacted. The engineering is the story, not the balances.
GOAT‑FARM
Autonomous options trader · live money
The premise
The model is not allowed to place an order.
The mechanism
A local 14B model forms the thesis on a 15 to 30 second loop. Tick-driven code reads it as plain flags and holds every order.
The rule
The brain can only make the system more conservative. It never sizes, never fires, never overrides a guardrail.
The proof
62,613 lines of Python trading a real account. Default state: flat and idle.
The live dashboard. A hypothesis book, a conviction gauge, and a decision ticker that prints every abort with a reason. Account identifiers and balances redacted.
It trades QQQ 0DTE options, single-leg long only, on a cash account. There's no paper mode
to hide behind. Account size is never hardcoded: every session starts from a live broker snapshot,
so sizing follows the account with zero code edits.
The guardrails are the actual product. Trade-count and budget locks, a daily-loss halt, an
equity-drawdown kill switch against a rolling high-water mark, a stale-data watchdog, and a global
exception handler whose only move is flatten, halt, alert. There is no
assignment code path at all, so a contract can never be carried into expiry.
Before deploy it ran an adversarial review with two named reviewer roles. Both blockers were
fixed first, and the residual risks are written down instead of omitted.
62,613 lines · 148 filesQQQ 0DTE, long onlynightly learning loop
JessAI
Autonomous SDR · four channels, all day
The failure
One prospect got 10 outbound messages in a day, 7 of them inside 11 minutes. 44 of 76 sends came from us, composing by hand with no counter running.
The mechanism
Code decides who gets messaged and when. The model only drafts what's said, and a draft that fails a validator is discarded, not sent.
The rule
Count what the recipient received, from every source. Four earlier caps failed because each sender counted only its own sends.
The proof
Proactive cap: 2 a day, 5 a week, enforced at the contact. Booked meetings on the board every week.
The wall: every live thread ranked the way a real SDR would work it. The "Held by the touch guard" lines are the point. Prospect names, numbers and companies redacted.
An always-on setter. It watches every inbound lead, decides who needs a touch and when, drafts
the message, validates it, sends it, and runs a deterministic post-booking playbook that no inbound
text can prompt-inject.
It carries an absolute honesty rule: asked sincerely whether she's a person, Jess says she's an
AI assistant, every time. The persona takes the sales pressure out of a first message. It never
exists to make someone think they're talking to a human.
The board keys one company to one prospect by email domain, so two people at the same pharmacy
never get worked as separate leads. Booking state only moves forward: a later sweep can never
un-book somebody.
These two were built months apart, for unrelated purposes, and landed on the same
architecture without either copying the other. In the setter, the model drafts a sentence and code
decides who receives it. In the trader, the model forms a thesis and code decides whether money moves.
Both state the constraint identically: the model can only make the system more
conservative, never less safe. That's a held principle showing up twice, not a trick reused.
Call Shadow
Live sales coaching · audio never leaves the machine
The failure
Sales coaching arrives after the deal is already lost. And every AI notetaker ships the audio to a vendor, which is a non-starter in healthcare.
The mechanism
Speech runs locally on the rep's own machine. Six discovery questions tick green as the prospect answers them, hands never touching the keyboard.
The rule
Quote a price before the six are answered and the whole screen goes red.
The proof
The coaching rules came out of 40-plus recorded calls. $49 a month, one double-click Windows install.
The console between calls. Six questions lock the price, the talk-ratio bar wants the prospect above 60%, and the right rail is the close ladder.
It also watches your talk ratio, arms a silence watchdog the moment you ask for money, and
writes complete notes to the right CRM contact when the call ends. A training room throws real
objections at you out loud and grades what you actually said.
$49 / month, shipped6 questions lock the priceprospect holds 60%+
The Hive
The agency cockpit · one place where the numbers are real
The failure
Every tool disagreed with every other tool, so nobody trusted any of them, so decisions got made on feel.
The mechanism
One cockpit: 95,766 lines, 145 API routes, 18 surfaces, and a 10,482-line finance library whose only job is making one number correct.
The rule
Render is not push. Every ad is born paused. Push is a human action, per client, every time.
The proof
9 of 9 labs ready, 225 of 225 ads through the pipeline, in the capture below.
The build surface. 9/9 fleets ready, 225/225 ads, and the push control reads "born PAUSED" in the corner. Client names replaced with numbered labels.
Creative production, Meta push, finance truth, fleet health, goals, performance and the Jess
board, all in one Next.js app that's load-bearing daily. The ad pipeline is a locked, documented
path: nobody writes a one-off driver script, and the last node before Meta is a human.
95,766 lines · 673 files145 API routes18 surfaces
The creative factory
Ads that grade themselves before a human looks
The failure
Image models misspell logos, duplicate words and leak prompt text into the art. At batch scale a human reviewer misses about half of it.
The mechanism
Three stacked gates. The last one is shown the rendered ad and the real logo file together, and returns a strict verdict.
The rule
A hard fail refires the ad through its own batch script automatically, looping until clean. The gate exits clean or not at all.
The proof
10,509 ad variants across 60 accounts. Zero shipped with a misspelled logo.
Client logos are composited, never generated. Every rendered string is quoted in the prompt,
because an unquoted line is an invitation for the model to invent its own words. And the gate is
deliberately lenient where strictness would be wrong: a false positive that blocks real work is
worse than no gate.
The finished output is in the output section: the ad wall and the lead
magnets on this page all came through these gates.
10,509 variants · 60 accounts3 stacked gatesauto-refire until clean
compound.buzz/preview
The pitch deck that lives at a URL
The failure
Prospects read decks alone, do the arithmetic against your price, and talk themselves out before the call.
The mechanism
17 cards at one URL, keyboard and swipe navigable, films playing inline, with per-card analytics and an ROI slider that uses their inputs.
The rule
No pricing on it, and never show net beside gross. If both are on screen they subtract, and the subtraction is your fee.
The proof
One link replaced the slide deck, the portal and the leave-behind. It's the only deck link a prospect ever gets.
Card 01 of 17. A break-even slide was built, then deliberately deleted: some slides lose deals by being honest in the wrong order.
17 cards · 744 linesfilms play inlineno pricing, by design
Land Scout
A land-deal machine that priced its own runs in pennies
36 mountain-town markets, each anchored to a real town and one plain-English rule implemented
server-side: about twenty minutes to a supermarket. Every hit was scored for under-market dollars
per acre against that area's own rolling history, because a per-acre number
means nothing against a national average.
A full sweep across all areas cost under a dollar, at roughly $0.0009 per listing. The rolling
ledger holds 2,206 listings, deduplicated so the same parcel never alerts twice.
Written in the past tense on purpose: the last run failed on all 35 areas it attempted, so the
qualifying counts are carried forward from earlier successful runs. The card gets its present tense
back when the run does. A file existing is not proof, including our own.
36 areas · 2,206 in the ledgerunder $1 a sweephonest about its last run
Coach OS
White-label SaaS · demo mode is a feature
A multi-tenant coaching platform for an external client, 12,417 lines, live on Vercel. The
client's database didn't exist yet and the demo couldn't wait, so with zero environment variables
it boots on seeded data and every screen works. Same components, one data-source switch. It demoed
real on day one and became production without a rewrite.
Authorization lives in Postgres row-level security, not in application code. The app never
decides access. Phosphor-terminal design language: one accent, mono for every number, command
palette on Cmd+K.
12,417 lines · 110 filesRLS is the boundaryboots with zero env vars
local-brain
The local brain
31 jobs across six phases, nightly, on my own GPU. Episodic memory, a concept graph, and a priming cache that gets session start under 100ms. Token cost: $0.00.
guardrails
The guardrail layer
Five gates that each make a specific past failure impossible to repeat. Every rail has a date on it, and the date is the day it cost something.
ghl-mcp
The CRM fleet server
A 33,065-line MCP server with about 285 tools, mounted ten times, once per client account. The difference between an agent that can read a CRM and one that can run a fleet.
finance-truth
The finance truth layer
10,482 lines whose only job is one correct number. Two processor keys once doubled revenue, and a half-loaded ledger once printed a -206% margin. Truth needs a definition, not a SUM.
myportal
The client portal
Co-branded dashboards with 180 days of daily history, refreshed nightly. Lead counts verified digit for digit against the ad platform's own reporting UI, because a dashboard that's slightly off poisons every number on it.
cinema
The video engine
14 pipeline versions. One reference build cut 10 ads from a client's real footage: 227 cuts, zero repeated shots. The logo is composited from the real file, never generated.
the-fill
The Fill
The whole funnel running unattended. Its best output was a diagnosis: 81.8% of booked calls showed up and 4.4% closed. Two numbers, one conclusion, and it pointed at the human on the call.
content-factories
The content factories
Ebooks and lifecycle email produced by pipeline under one standard: a hard content gate runs before any image budget is spent, so a bad draft can't cost money.
call-stack
The call stack
Everything in the 48 hours around a sales call: reminders in the prospect's real timezone, a watch signal to the ad platform, transcript capture after. Rule learned live: a pre-call text never pre-qualifies.
close-school
Close School
An offline-first sales training PWA with audio lessons, built to work on a flight. A phone can't play media from a file path, so it had to be a real installed app.
jess-cockpit
The Jess cockpit
The 4,170-line wall the setter is judged on: four outbound systems side by side rather than merged, because they fail differently and a merged view hides which one is broken.
the-long-tail
The long tail
Twelve small apps that stayed small: a deal desk, a review app, a pricing calculator, a launch page. Most could've been a platform. None should have been.
04 · Campaigns and CRM
The creative is what people look at. The funnel underneath it is what gets measured.
Every account runs to a fixed funnel mix. The distribution is set in advance, the offer stack comes
from that client's own verified strengths, and each stage gets its own creative treatment, because a
scroll-stopper and a closing offer are not the same photograph.
60%
Top of funnel
Awareness. Carries the audience tag in the picture, unmissable: "for Miami GPs",
"for full-arch dentists". If a viewer cannot tell in a third of a second that the ad is for them,
they scroll. Documentary treatment, because polish reads as an ad.
30%
Middle of funnel
Consideration. Pillar-specific and territory-specific. Clinical product photography and macro
detail, because precision at this stage has to be visible in the photograph itself. No audience tag.
10%
Bottom of funnel
The direct offer. A warm face plus a credibility object. This is the only stage where the number
leads, and only when the number genuinely is the outcome.
40%
Presell inventory
The free tool, the buyer's guide, the review builder. Real utility that earns the click before
anything is asked for, produced by the research pipeline so every claim traces to a source.
CRM architecture
23 isolated instances, one deployable standard
Every client gets their own CRM instance built from a versioned snapshot: pipelines, stages,
workflows, custom fields and values, calendars. One standard layout, deployed by slug, which turns
standing up a new account into a build script.
Lead intake→Case production→Retention→Referral
Stages advance on tag or on time, so the pipeline reflects reality without
anyone dragging cards. Retention and referral are stages in the pipeline, not an
afterthought, which is what makes reactivation systematic instead of a quarterly panic.
Guardrails that matter
Why automating a live CRM is safe
Every build runs a pre-flight first: confirm the account exists, confirm the credential can reach
it, and read that client's specific restrictions. Only then does a build start.
Then the hard rules. Never rename a pipeline on a live account, because a rename orphans every
open opportunity. Never delete a stage that has deals in it. Never skip capturing stage IDs, because
workflows reference them and rebuilding re-rolls them.
Automation against production data is only worth having if it refuses to do the
thing that quietly destroys a quarter of pipeline.
Lifecycle and re-engagement
36 emails, 12 months, one narrative
A full-year nurture sequence per brand. It reads as one continuous story, with branch points for
people who go quiet and separate recovery paths when something breaks.
The conversation AI picks up from there inside the CRM, so a reply never lands in an inbox nobody
is watching.
Self-promo, run like a client
The agency is its own hardest account
Buzzword's own acquisition runs through the identical pipeline: the same funnel mix, the same
creative gates, the same CRM standard. The ads on this page are ours, not a client's.
Running our own demand generation on the system we sell is how we find out what breaks before a
client does. Every improvement gets tested here first.
05 · The output
1,823 finished ad creatives. 268 lead magnets. 2,091 videos. Same eight people.
A delivery system is worth whatever it ships. These are finished, shipped assets, produced
through the pipeline with brand rules, compliance gates and an automated QA pass applied before
anything reached a client.
Lead magnets and long-form. Researched,
written, illustrated and typeset by the Factory, then gated for claim safety. Every factual claim traces
to a source that was actually fetched and read.
06 · What I actually run
A founder owns every function. I have run all of them with a P&L attached.
01
Marketing
Positioning, messaging, paid social, lifecycle, content and creative strategy across 21 accounts and four brands. $800K+ in annual media under management.
02
Sales
Built the outbound motion and the offer, and closed the deals myself. Regional exclusivity across roughly 100 US territories. Previously added $500K+ to pipeline in under three months as an SDR at Carbon.
03
Management
Hired and lead eight across account management, paid media, CRM engineering and social. Set the operating cadence and the quality bar, and decided what escalates to me.
04
Finance
Own the P&L, pricing, margin and forecasting. Built the finance cockpit that reconciles revenue, spend, payroll and churn without a bookkeeper in the loop.
05
Operations
Designed the delivery system itself: intake, production, QA, approval, reporting. The reason 21 accounts run on eight people.
06
Engineering
Wrote the platform. Node, Python, Next.js, Postgres, local and hosted LLMs, MCP servers, browser automation and evaluation harnesses.
07 · Track record
Product marketing at scale, then proving it on my own balance sheet.
Jan 2025 to present
Founder, CEO
Buzzword Strategies LLC
Bootstrapped to $1M+ ARR in 19 months with no outside capital.
Built and operate the multi-agent AI platform that runs client delivery end to end.
Hired and lead a distributed team of 8 across four functions.
Direct $800K+ in annual paid media with automated pacing and reporting.
Led go-to-market for 10+ launches across materials, software, hardware and automation.
Exceeded launch targets by over 200%, driving month-over-month adoption.
Built pricing models with finance and product, balancing adoption against margin.
Created the enablement suite used across 100+ customer interactions.
May 2022 to Jan 2023
Senior Sales Development Representative
Carbon, Inc.
Built outreach frameworks from user pain points, adding $500K+ to pipeline in under three months.
Prototyped internal AI-assisted tooling to score lead quality and tighten the marketing-to-sales handoff.
Nov 2020 to Apr 2022
Senior Technical Product Manager
Partners Dental Studio
Managed six new product lines in a regulated healthcare environment.
Built operational workflows and automation that raised throughput and cut error rates.
Available now
Looking for the role where the system is the product.
VP of AI, Head of AI, AI in marketing, or a CMO seat at a company that wants an operator who ships.
Fully remote. Buzzword is built to be sold and the platform runs delivery without me day to day,
which is exactly why I'm looking.
