AI Transformation
I run businesses on AI.
THE RANGE
Every function of a business now runs on AI, from research through media to the code inside a shipping product. Here are 26 systems I run across 8 lanes, each doing one concrete job and wired into tools you already use.
These functions run in production today, and every card names its wiring so you can check the work yourself.
Infrastructure & Models
I know the whole landscape and pick per job. Every option here carries a judgment: when it wins, and what it costs you to run it. Four layers stack up, from the runtime the agent lives in to the glue that connects everything, and I can defend the call at each one. A single-vendor stack is the seller's convenience, rarely the build's.
The systems I run today are built this way. Cheap models carry the volume, frontier models take the hard reasoning, open weights run wherever sovereignty or cost demands it.
Open, frontier, or both
Two axes decide it: how hard the reasoning in the step is, and how much volume runs through it. Most real builds land on the diagonal between them.
TAP A REGIONSee it run
Press play. One real research operation runs end to end — the machine rips the site and its APIs, finds the decision-maker and verifies their contact, pulls the brand’s design system, then reconciles every source to catch the contradictions and rank the follow-ups worth chasing.
Nothing here is staged. The tools are the real ones, the modules are the real machine, and the figures — 35+ files, nine contradictions, thirteen leads — come from an actual run. It ends in a full audit deck you can open below. Only the target is anonymized.
Built For
Six businesses that look nothing alike. One operator ran the AI in every one.
Spree FinanceFintech
KuCoinConsumer exchange
EntangleMarketing infrastructure
OlmiRegulated EU fintech
QuraniumQuantum-resistant L1HOW IT STAYS TRUSTWORTHY
Before any output reaches a business, it crosses gates I built into the system. The operating layer that runs the agents also governs them: a verified canon feeds them, adversarial readers check what they write, and every irreversible call waits for me.
BlocksAn ungoverned AI invents facts and states them with full confidence, so a made-up claim reaches your customer as truth.
MechanismNo agent may state what the canon can't back, and every claim in the canon carries a confidence tier. That canon is verified research held on pgvector, across 10 company domains.
On failureA claim the canon can't back gets flagged UNVERIFIED, and no agent may treat the claim as fact.
BlocksAn ungoverned AI writes in tell-tale machine voice and can run a destructive command on live systems with nothing watching.
MechanismAs the draft is written, enforcement hooks fire in place: a prose linter blocks AI-tell writing, a command-safety guard stops dangerous shell calls, and a model-policy guard holds each agent to a pinned model. I ship this kind of guard in the open too, from a self-built MCP gateway to an ads-transparency scraper I forked and republished as an MCP server.
On failureA flagged write gets blocked before it lands, so nothing off-policy leaves the agent that wrote it.
BlocksAn ungoverned AI ships a draft that confuses its reader, who finishes it unsure of the claim and unsure what's being asked.
MechanismIndependent reader agents get the finished draft with zero context, the way the recipient will, and each restates the claim and the ask. A final preflight then reads the rendered draft, so the check runs on what ships.
On failureWhen the readers converge on confusion, or a condescension flag trips, the draft dies and gets recomposed.
BlocksAn agent that can call paid APIs can burn a budget overnight.
MechanismEvery paid model or media call is wrapped in a gate that prints its dry-run cost and waits for an explicit go.
On failureNo go, no call: the job parks and reports what it wanted to spend.
BlocksAn ungoverned AI acts on guessed contact data and fires messages at unverified addresses.
MechanismEvery address must be observed at a real source, then checked against the mail server to confirm the mailbox exists, before the system runs a send.
On failureAn address that fails a check stops the send, and the gate fails closed rather than leaking an unverified contact out.
BlocksThe fear with any AI operation is that it runs end to end unsupervised and makes the calls a business can't take back.
MechanismDetection, drafting, formatting, and publishing run automated. The judgment stays with me: what ships, what money moves, what gets promised, and how a crisis gets answered.
On failureThis is the only gate with no kill path, because it's a standing doctrine: the automation stops here and I take these calls myself.
With the safety built in, the open question is what the system is worth.
THE MATH
This board prices the real cost of running AI for an organization, three ways: frontier vendor APIs and plans, a hybrid of frontier judgment and open-weight bulk, and open or self-hosted weights. Each sits against the manual baseline. Every figure is sourced, every assumption labeled.
I run my own operation on this same math. My spend gates print a dry-run cost before any paid call fires, my video work bills against the self-hosted GPU receipts pinned here, and I source across the same three models. No number here is invented.
- API rates read input / output per 1M tokens.
- Plan seats read annual / monthly per seat.
- Video reads per second; GPU reads per hour.
- DERIVED = a monthly total computed from a captured volume × an annex rate.
ONE SEAT: FLAT PLAN VS METERED API
A seat is one person working with a frontier AI vendor. You can pay a flat monthly subscription, or pay per use, priced in tokens: small fragments of words. With light use, paying per token is cheaper. Each orange square marks the point where both cost the same; past it, the flat subscription wins. The math assumes three input tokens for every output token.
CLAUDE SEAT- PRO $20/MO: cheaper than metered SONNET 5 past ≈5M tokens/mo
- MAX FROM $100/MO: cheaper than metered SONNET 5 past ≈25M tokens/mo
OPENAI SEAT- PLUS $20/MO: cheaper than metered CODEX past ≈4.2M tokens/mo
- PRO $100/MO: cheaper than metered CODEX past ≈20.8M tokens/mo
- PRO $200/MO: cheaper than metered CODEX past ≈41.6M tokens/mo
No vendor publishes a break-even, so this one is ours and labeled.
Neither vendor publishes numeric usage caps, so this chart shows and implies none.
Drawn as a shape rather than measured data: the manual cost of this work stays flat month after month, while each automated run costs close to nothing once the system is built.
A small Spree team operating at 30-person throughput.FAQ
The doubts a careful buyer still has, answered straight, each with a pointer to where the page proves it.
0 / 7 ANSWEREDMost fail on adoption and governance long before model choice matters. A clean, verified canon beats a bigger fleet of agents.
Governance and adoption matter more than model choice.
An agent that runs real work unattended against live systems, reading a verified canon and clearing gates before anything ships. That bar sits well above a chatbot answering questions in a window.
A production agent ships real work unattended, under gates.
→ §4The judgment: what ships, what money moves, what gets promised. Detection, drafting, and publishing run automated underneath.
Every judgment call stays with a human.
→ §6Usually both. Frontier models earn their price on the hard reasoning while open or self-hosted weights carry the high-volume work, so most real builds land on a hybrid.
Most real builds are hybrid, chosen per job.
→ §3→ §7Every agent reads from a verified canon where each fact carries a confidence tier, so it answers from checked ground. Adversarial readers with no context restate the output, and nothing ships until it clears the gates.
A verified canon and adversarial gates keep agents from making things up.
→ §6Models are swappable parts. Each agent runs a model I pinned for a stated reason, so when a better one ships I slot it in without rebuilding the system around it.
Models are pinned, swappable parts, each chosen for a stated reason.
→ §3No. The same architecture runs research, code, video, data, and hiring screens; any function that sits on a verified canon and gated agents works the same way.
The same architecture runs any business function.
→ §2