AI systems and automation for fast merchant onboarding

Adyen is converted at the ECB 2025 average of 1.12998 USD per EUR and Square is the sum of its four published quarters, so each bar follows its own company's volume definition.

The build

• Onboarding and pricing pages carry dated figures and onboarding times, so an assistant can quote them and a merchant can check them.
• Schema markup and plain-text documentation make the same pages readable to an assistant in one pass.
• Page content comes from complaint mining of public sources, where 85 of my 2,017 hand-labelled complaints concern holds and payouts.

Agent checkout protocols

• The Agentic Commerce Protocol, announced 2025-09-29, lets an assistant complete checkout on a buyer's behalf.
• Google AP2 was announced 2025-09-16 with more than 60 organizations behind it.
• Visa and Mastercard define agent identity and tokens on the card rails, so the protocols settle on the networks MiCamp already uses.

Holds and payout timing

Funds are held following approval and payout schedules change without notice, so the first fix is a published, dated schedule.

Fee transparency

Rate changes and itemised fees differ from the quoted price, so the quote and the statement have to show the same lines.

Support response

Replies are scripted and escalation is slow on integration questions, so an integration question needs a route to someone who owns the integration.

Integration fit

Some processors do not connect to the software the business already runs, so the connector belongs inside the onboarding and not at the end of it.

Termination without a route back

Account closures come with no stated path to reinstatement, so a closed merchant needs a written route back.

hand-labelled public complaints sit in my mining pipeline, where I rank them by count and severity to decide what the onboarding pages answer first.

Published network limits

• Visa VAMP merchant excessive was 220 bps and fell to 150 bps on 2026-04-01, with at least 1,500 monthly events.
• Visa VAMP acquirer is above standard at 50 bps and excessive at 70 bps, so the acquirer's ratio moves with its merchants' ratios.
• Mastercard's excessive chargeback program starts at 100 chargebacks and 1.50%, with fines from $1,000 to $200,000.
• Visa Order Insight let merchants deflect 40 to 45% of confirmed first-party disputes.

Controls

• Underwriting screens site, product and licence, and runs rules in shadow mode against historical applications before release.
• Rolling reserves and delayed settlement are sized to the ratio, so exposure follows the risk.
• Continuous chargeback scoring and checkout-page script monitoring (PCI DSS 4.0.1, 6.4.3 and 11.6.1) catch movement before a threshold does.
• Dispute evidence is assembled automatically and every automated decision is logged for replay (Federal Reserve SR 26-2).

Merchant onboarding automation

A setup questionnaire configures the product to a business at sign-up, and a readiness checklist and restricted-terms gate screen high-risk merchants before an application goes out, so a weak application is caught before it costs a review.

Processor choice screen

It compares 4 processors with every fact carrying a source and a checked date, and offers three ways to connect, an embedded account, the business's own processor or a referral, so the merchant picks the route that fits.

Complaint mining

I labelled 2,017 public merchant complaints by hand and found 85 on holds and payouts, the cluster that decides which onboarding pages and support flows come first.

Sandbox integration testing

I wrote a shop's card path against a processor's sandbox, broke it on purpose four ways with all four caught, and caught 17 of 17 planted faults in an offline queue.

Chargeback and compliance monitor

It scores a merchant's chargeback ratio against card-network limits, including Visa's 150 bps threshold from April 1, so a high-risk account is flagged ahead of the limit.

Medicus

I founded a state-licensed company at 24, raised investment and secured one of the first licenses issued in Oklahoma in 2018, then sold its products myself to 75+ stores for $1.3 million of revenue in its first year. I hired over 20 people, set salaries and raises, built a budget for each part of the business and wrote the SOPs and training the team ran from.

Volta

I commissioned a $500,000 cryogenic extraction plant three prior hires had failed to run, diagnosed the fault from the process behavior, and re-routed the system with $5,000 of parts, which took a business with two years of no sales to $60,000 a month within two months. I recovered batches already written off, 50 kg of extract to 30 kg at 99% purity, by building a pesticide remediation line.

Autonomous fix loops

I run research, build, test and repair loops on predetermined methods that do not drift, watched by over 750 daily checks and ratchets that only let counts improve, which closed 3,531 requests in 30 days.

Per-pipeline memory

I built a memory that saves what each loop learns into its own pipeline folder, 14 folders and 703 loop records, so parallel sessions never overwrite each other or mix one pipeline's work with another's.

Custom model stacks

I train and score custom models, 477 across 11 Kaggle competitions, using neural networks, distillation, learned combiners and stacked ensembles to build a stack for each task.

Retrieval and research loops

I combine local indexes over 82,859 passages with research loops that scan hundreds to thousands of sites a day, including about 17,900 reviews and posts, and pull the newest tools and methods into the system.

Agent orchestration

I run parallel agents in isolated clones with enforced file ownership and a single merge, behind 66 gates that refuse unsafe writes and unproven claims of done.

Token optimization at scale

Forty enforced methods, among them local models, batching, slicing, background tasks and indexing, make Claude use 30 to 50 percent fewer tokens on identical graded work: 50 percent on a 14-task benchmark and 30 percent on an 84-task one against a default agent.

Closed-loop process control

I tuned soak times and temperatures per batch against HPLC assays taken before and after extraction, the measure-and-adjust loop I now run on software.

Human-in-the-loop change

I built the agent a shop owner types requests into, where it edits a disposable branch, a reviewer battery checks it in a sandbox and a human approves, and 29 requests have passed.

Point of sale

Counter, split tender, stock and reports in one retail system, with 31 pages, 53 API endpoints, server-recomputed tax and a spreadsheet sync that updates a sale instead of duplicating it.

Processor research

I compared 30 payment processors on methods, onboarding, restricted categories, reserves and holds, and recorded a cell as Yes only where the vendor's own page said so, so every yes traces to a source.

Reconciliation

I built a tender ledger for split payments and settlement matching that ties each deposit to the sales it covers, and cleared 25 of 25 payment-risk cases on holds, reserves and chargeback limits.

Integration layer

I built a commerce platform's connectors for Stripe, Shopify, WooCommerce, TikTok, Twilio and a card processor's gateway, checked by 192 Stripe tests and 82 webhook-signature tests, so each connector is proven before a merchant relies on it.