Engineering for brands whose products carry regulated claims
One senior engineer, directing a team of AI agents inside your business.
The agents produce the work your revenue depends on: marketing, web development and research. I stay through live use. Your people keep control. Your business keeps the system.
The team is software; the accountability is one person. I brief the agents in sessions I run, verify their work against each other and merge what survives. Nothing acts unattended.
From one real inspection · 2026
The shop said 5,490 customers could be mailed. The email platform could reach about 150. Nothing anywhere said a word.
One of three findings published here, reproduced before reporting. No campaign has been sent to any of those records, and no customer outcome is claimed.
One named builder: David Bentley · Built in your accounts, not mine · Nothing changes state or sends without your named approver · A documented handover
Evidence you can reproduce
A recent inspection found the following.
One private engagement with a small independent British supplement business, inspected in 2026. Six findings were reported privately, each reproduced before it was reported; three are published here. The business is not named and never will be: a client’s private weakness is not my marketing asset.
Every tool can say “working” while the workflow is broken. The failure lives in the join nobody owns.
Five of the six findings sat between two systems, in a join nobody owned: storefront to email platform, sender to DNS, page to visitor, form to server, staging to live.
Three findings · confidence: confirmed · each reproduced before it was reported
The exception · not a join
Finding 1 · Compliance
A claim the product was not entitled to make.
Severity: high
A product named after a mineral carried a claim for that mineral at 13% of the Nutrient Reference Value. The GB register permits a claim only at 15% or above. The product named after the ingredient could not lawfully claim it.
The join · storefront to email platform
Finding 2 · Data
An audience written off by one account action.
Severity: high
A single account-wide suppression event in April 2025 was mistaken for individual opt-outs. The storefront’s marketing flag said 5,490 people could be mailed; the email platform could reach about 150. Neither system flagged the gap, because each is accountable only for its own half. Restoring a profile inside the platform is not the same as having a lawful basis to email it, and no campaign has been sent to any of these records.
The join · sender to DNS
Finding 3 · Email
Email authentication broken for months.
Severity: high
SPF, DKIM and DMARC were misconfigured: the sending platform was authenticating against the registrar while the DNS actually lived at Cloudflare. Every marketing email was signed as the platform, not the brand.
None of these needed AI to exist. All of them needed someone to look, check what “working” means and write down how to reproduce what they found. The findings were found by hand; what I build from them is where the agent teams work. Every finding I report comes with its reproduction path and a plain statement of what I could not check.
Selected work
Three systems built for a €3 million aviation training business
Sofema Aviation Services trains engineers under EASA regulation, where a wrong answer carries a regulatory cost. All three systems below were built agent-assisted and human-directed, and two are publicly reachable today.
Where this evidence stops. I did not win this work in open competition, and no independent buyer weighed me against anyone else and chose to pay. So it shows what I can build inside a regulated operation. It does not show that the market has chosen me, and I would rather say so here than have you assume otherwise.
Training ROI Calculator
Publicly reachableIn testing
Shows the full cost of classroom training against Sofema’s Corporate Freedom Pass.
Course Finder
Publicly reachableIn testing
Uses four questions to route a learner to relevant courses, bundles and diplomas.
Cited EASA answer system
Working previewNot deployed
239 plain-English regulatory pages, each tied back to source law, with search and review by Sofema Aviation Services. No deployment decision has been made, and it may never ship.
Capability evidence only. No adoption, revenue or return-on-investment result is claimed.
How the work runs
A team of agents. One accountable builder.
A team of agents changes what one engineer can deliver, not who answers for it. You work directly with David Bentley, the engineer responsible for the work: there is no sales-to-delivery hand-off and no junior delivery layer. The agents are software; judgement does not delegate.
The engagement, end to end
Live use finds the next work
Find the work worth changing. I map what happens now, who owns it, where it fails and what result would make the work worth doing.
Build inside the real environment. Agents produce code and drafts under my direction, verified against the data, tools, permissions and exceptions the workflow actually uses.
Prove it before giving it authority. I agree tests with you, run the system on representative work, read the agents’ work against each other and keep a person in control of consequential decisions.
Operate, document and hand over. I stay through live use. Your people receive the source, operating records, dependencies and instructions needed to keep control.
Access is read-only by default, on credentials you issue, in your own admin, revocable by you at any time.
For a live engagement, agreed checks report the current state. When something should change, the system proposes the change in writing, with the evidence and prior state captured.
The delivery record captures agreed checks, proposals, approvals and changes. A quiet month is reported as a quiet month, not left blank.
This is slower than full automation, deliberately. In a regulated business, the expensive failures are the confident, unattended ones. The expensive waste is quieter still: a system that works perfectly and automates a step nobody was waiting on. That is what stage one exists to prevent.
It stays with your organisation.
The source, data, workflow definitions, keys, logs and documentation live in accounts your organisation controls. Every external dependency is named. If the engagement stops, the system and a tested handover stay with you. The underlying AI model may still be supplied by a third party; where that is true, I say so and document the exit path.
There is one of me. Here is what that means.
HigherSelf Systems is one person, and you should know what that means before you buy. From month one you hold the documentation, regularly exported logs and a runbook any competent engineer can pick up. Holidays are scheduled and told to you in advance. If I am ill, you hear it from me, not from silence, and anything urgent routes to your named person, not to a pager I might miss.
The agents are the capacity; one person is the accountability.
Engagement and fit
The engagement, in the open
Each step is complete in itself and refusable, and the next one is only worth buying if the previous one produced evidence in your own numbers. The first market is deliberately narrow: UK supplement and wellness brands turning over £2 million to £20 million, with their own online store, Klaviyo for email, and a credible human coach or subject expert. If that is not you, or you are not on Klaviyo, write anyway and I will tell you honestly whether I can help.
“Regulated” on this page means any business where a wrong automated action carries a legal or safety cost, not only aviation: a public register decides what a supplement label may say, and finding 2 above is what it looks like when nobody checks.
Free · one page per organisation
One page, read properly
Reply
Within two working days.
What you get
One public page of your site read by hand, by me: the served markup, the forms (read, never submitted), outbound links, HTTP headers and the domain’s public email authentication. The reply names either a specific problem I verified, with the exact steps to reproduce it in your own hands, or what I checked, what was clean and what I could not see from outside.
Exit
One page, one reply, no sequence. A clean answer is the deliverable, not a consolation.
Fixed fee, in writing before any work · 15 working days
The Klaviyo Audience Recovery Sprint
Before any fee
A short screen share inside your own Klaviyo account. I proceed only if all four conditions hold: a material account-initiated suppression event; enough consent, purchase and checkout history to identify at least one credible contact-basis cohort; your written ownership of the lawful-basis decision under PECR; and a group large enough to justify the fee in your own historic order values and campaign figures. If any condition fails, I decline the work. I do not produce a report from this check; it decides whether the work proceeds.
What you get
Execution, not an audit, inside your own account: a dated baseline before any change, an audience ledger recording why every record sits where it does, permanent protection for genuine unsubscribes, bounces and complaints, restoration only of the records the evidence and your approval support, an authenticated sending domain, and one measured campaign your named owner approves before it goes. The analysis, drafts and campaign build are agent-produced in sessions I run, verified and merged before your named owner sees them.
What proves it worked
A written results account seven days after the send, in your own numbers, ending in a go or no-go recommendation. No tracking pixels, and open rate is not used as evidence.
Exit
Complete in itself. Start the build within 30 days and the Sprint fee is credited against stage one. There is no other discount and no performance fee.
Boundary
Restoring a profile in Klaviyo does not create consent. Before any send, your named owner documents the lawful marketing basis for every group mailed, and records without one stay out.
Three separately approved stages · 90 days
The Catalogue-to-Practice Build
What you get
One bounded web-and-Klaviyo build for one product line: a goals-and-preferences assessment, a personal plan with a named human review step, a paid monthly coaching offer, and the journeys to test, sell, onboard and retain it: one product line turned into a guided monthly relationship for the customers your brand serves. In supervised sessions, agents produce the research, drafts, web components and campaign builds; I verify and merge the result in your accounts.
What proves it worked
Stage one proves demand with a real, approved campaign before anything expensive is built; a weak response stops the build at that gate rather than being relabelled as engagement. The final deliverable is the installed, working route and a day-90 decision on whether ongoing operation is worth buying. Revenue becomes evidence only when your customers act and the result is measured.
Exit
You can stop at every gate. Work from an approved stage is complete and payable; nothing rolls forward quietly. Stop after stage one and you keep the tested offer and route, the baseline and threshold, the approved campaign, its written results account and the go or no-go decision.
Boundary
I have built this route once, end to end, for the business in the findings above. It is built and it has not yet run: no campaign has reached a real customer and no result exists. You are buying the build, gated so a weak response stops it, not a proven playbook. The promise is the installed route, not a revenue figure.
Monthly · five named delivery days · six-month minimum, then quarterly
Embedded Revenue Operation
What you get
One senior engineer and a directed agent team operate and improve the system built during the first 90 days: one active customer programme, one prioritised improvement at a time, a weekly approval rhythm, agent-produced analysis and campaign construction, and a monthly commercial account in your own numbers. I verify and merge the agents’ work before your approver sees a proposal.
What proves it worked
The monthly account reports the agreed operating checks and commercial measures in your own numbers. Monitoring is named only after specific checks have actually been built, tested and put into operation for your business; it is never assumed.
Exit
Six-month minimum, then quarterly review. You choose whether to continue at each review.
Boundary
Not an autonomous marketing department, not continuous monitoring, not a 24-hour response. The agents work only in sessions I direct; nothing monitors, decides or sends by itself.
What is not included
The build does not include blood testing, clinical or medical advice, social media, paid advertising, a general website redesign or unlimited campaign production. No revenue, conversion or deliverability outcome is promised. The always-on agent I am working toward is a specification, not running software, and it is not part of anything sold on this page.
Show me the work that keeps getting stuck.
What happens next
You write: one sentence about the workflow, or one public page URL from a work email at the same domain.
I reply: within one working day for a workflow, two for a one-page read.
If the workflow fits, we screen-share inside your own account; a one-page read ends with its reply.
I refuse the work, or we agree the price, the boundary and the named approver, in writing. Only then does paid work, or any access, begin.
Write to david@higherself.ai with one sentence about the workflow. If you know them, add the person responsible and what happens when it fails; if not, send the sentence alone and I will ask one question rather than a brief. If it does not fit, I will say so in the first exchange and point you somewhere better if I can.
If you would rather not write to a stranger first: one public page of your site, read by hand, is free.