Brian Palma AI Automation Consultant

I build production automation systems that run real business operations.

Self-hosted n8n, LLM APIs, and the engineering around them: research engines, governed approval workflows, content studios and knowledge pipelines that keep running when nobody is watching them.

For agencies staffing automation roles, and for businesses sitting on a process that eats hours every week.

Book a scoping call Pricing Capabilities Deck (PDF) Resume (PDF) LinkedIn
CEO Office retainer
A global expansion and EOR services provider operating in 145+ countries
Operations builds
North American field service businesses: intake, CRM and reporting
Own venture
MeisterSmith, an e-commerce content engine I run end to end
Practice

What I do, and who I do it for

I am an AI automation consultant, and I have ongoing engagements with multiple clients. The systems run on self-hosted n8n with LLM APIs behind them. For a global expansion and EOR services provider I hold a CEO Office retainer, building for sales enablement, the central PMO and the people function. For North American field service businesses the work runs through intake, CRM and operations reporting. For my own e-commerce content venture I own the entire pipeline, from product research to published video. Three different rooms, one pattern: find the manual work, write a proposal that names the deliverable and the risks before a single node exists, ship it through a governed intake and review process, then hand over a system the client owns and can run without me.

A note on the work shown here.
Client engagements are covered by confidentiality. Every system below is described by what it does, how it is built and what it changed, never by client data or client name. Where a figure is an estimate, it says so. Where a system has been validated rather than run at volume, it says that too.
The numbers

The practice, counted

Read from delivered work, not from a proposal. Where a figure belongs to one engagement, the label says so.

25+
Systems built across my own platform and client engagements, from content engines to research and approval systems
50+
Reusable automation skills built for my own platform and client tooling
10+
Specialised AI agents running across my platform and client builds, each with its own scope and tools
160+
Python automation tools authored across those systems, covered by 130+ test files
290+
Workflow nodes authored inside a single client engagement. Largest single workflow: 106
25+
Tracked releases across client systems, with changelogs and rollback checkpoints
38+
Adversarial and regression test scenarios executed across client and platform systems
20+
Business systems integrated across CRM, M365, ticketing, email, messaging and e-commerce

System, skill, agent and tool counts are read directly from my own working repository. Client node, release and test counts are measured and rounded to the nearest unit of work. Any time saved per task is an estimate and is labelled as an estimate wherever it appears on this page.

How I work

Fixed price, tested on your data, yours at handover

Seven steps, the same every time. You know the price before the build starts and you sign off before the final payment.

1

Scoping call, free

We walk the process end to end. I come back with what is worth automating, what is not, and what it would take. No charge, and no obligation to go further.

2

Fixed-price proposal

Named deliverables, the architecture, scope split into must, should, could and will not have, the risks, and a phased plan. One price, agreed before anything is built.

3

Kickoff

Access, credentials and a single point of contact. Credentials belong to the project, not to me, and access stays read-only unless the build genuinely needs to write.

4

MVP in front of you early

The smallest version that does the real job, delivered early rather than at the end. You correct the direction while correcting it is still cheap.

5

UAT on your real data

You test the system on live data before the final payment. If it does not pass, it is not done, and I keep working until it does.

6

Documented handover

Workflow source, written documentation and a walkthrough with whoever will own it. You keep the workflows, the credentials and the docs. Nothing is locked to me.

7

Two weeks of support, then optional retainer

Two weeks of post-handover support is included in the price. A monthly retainer after that is entirely optional, and only makes sense if you want ongoing changes.

What this means for you

The risk sits on my side of the table

The price is known before the build starts, so there is no meter running. You test the system on your own data before the final payment, so you are never paying for something unproven. Everything is documented and handed over, so you own it outright. If you want me gone the day support ends, it all still runs.

Selected systems

Not a chatbot bolted onto a business. Systems that carry real work.

Each one shown as problem, approach and result, with the build detail and the architecture behind it. Status is stated plainly: live, in pilot, or built and validated end to end.

01

Company research engine

Live, in sales pilot
Problem
Account research was manual, so only marquee accounts ever got a brief. Everyone else went into the call cold.
Approach
One self-service workflow behind a branded web form, running two modes: an external prospect brief and an internal account brief.
Result
A finished, source-attributed PDF brief by email in minutes. Accuracy iterated to 9 out of 10 against CRM ground truth. Coverage moved from marquee accounts only to any account on request, replacing an estimated 30+ minutes of manual research per account.
Build: 106 nodes, 20 tracked releases. The prospect mode combines live web research, commercial enrichment and CRM footprint. The account mode builds from CRM data: deal map, buying committee, engagement recency and tailored talk tracks. A three-tier deduplication cache means a repeat request costs nothing. Allowlist validation runs on every link, so an invented URL cannot reach the output.
Web formcompany name
→
Mode routerprospect or account
→
Research + enrichweb, commercial, CRM
→
Allowlist + cachelink validation
→
PDF briefemailed, attributed
02

Governed approval system

Built and validated end to end
Problem
Hiring, offer, promotion and termination requests moved by email. No single record, no reliable audit trail, no way to see where a request was stuck.
Approach
One workflow serving four branded intake forms, with a four-level approval chain on tokenised single-use links and a governance workbook as the system of record.
Result
Validated end to end, with 12 guard rails confirmed to refuse correctly. Adding a fifth request type is a data change, not new nodes.
Build: 65 nodes, schema driven. A single Python schema generates the forms, the workbook layouts (38 to 46 columns depending on request type), the approver summaries and the expiry sweep. Approvers can approve, decline, or request more information. A bounce-back loop returns the request to the gate that asked for it while keeping earlier approvals intact, so nobody re-approves work they already signed off.
4 intake formsbranded, validated
→
Schemaone Python source
→
4-level chainsingle-use tokens
→
Bounce-backrequest more info
→
Governance workbooksystem of record
03

Social content studio

Live
Problem
Executive social content needed several formats and several voices, produced quickly, without a model inventing a link or drifting past a length limit.
Approach
A branded form producing text posts in multiple executive voices, designed 1080x1080 branded graphics, and carousel PDFs.
Result
Live. Stress tested across roughly 26 scenarios including prompt injection, with no system prompt leakage.
Build: 48 nodes. The model picks one of four hand-built HTML templates, so the graphics carry real type rather than generated pixels. Carousels are capped at ten pages by construction. A deterministic punctuation sanitiser and sentence-count enforcement run after generation, because a character limit does not bind a model but deleting a sentence does. The link allowlist is enforced twice: once at generation, once at render.
Brief formtopic, voice, format
→
Generateprose only
→
Sanitise + enforcesentences, allowlist
→
RenderHTML templates
→
Post, graphic, carousel3 formats out
04

Meeting knowledge pipeline

Built and validated end to end
Problem
Meeting transcripts either vanished after the call or got archived without the organiser ever agreeing to it.
Approach
Capture ended group meetings that have transcripts, classify them, ask the organiser for permission, and only archive and publish on an explicit yes.
Result
Built and validated end to end. The permission gate was confirmed to hold: nothing is archived or published without an explicit yes from the organiser.
Build: 37 nodes, private by default. The cost architecture matters here: when the meeting platform already wrote its own recap, the pipeline reuses that recap verbatim and spends zero generation tokens. A model is called only for a single topic tag. On approval the transcript is archived to a topic and year folder and a lightweight summary is published.
Meeting endstranscript present
→
Classifyone tag call
→
Permission gateask the organiser
→
Archivetopic and year
→
Publish summaryrecap reused
05

Revenue reporting, through CTO review

Built and validated end to end
Problem
A five-tier automated revenue reporting design carried real editorial risk, a wide audience blast radius and open questions about credential governance.
Approach
Took the design through CTO review and architected in all nine required revisions rather than arguing them down.
Result
Conditionally approved with nine required revisions, all nine built in. Built and validated end to end, with a four-cycle calibration phase and a signed exit criterion designed into the release path.
Build: isolated per-tier sub-flows, so a partial failure still ships a partial set instead of nothing. Recipient list fingerprinting to catch an audience change before it sends. A four-cycle calibration phase with a signed exit criterion, so nobody declares it accurate by feel. Credential governance with named owners and rotation triggers. Run-time self-export, so the workflow is always reconstructable from its own output.
Source datafive tiers
→
Isolated sub-flowspartial failure safe
→
Recipient fingerprintblast radius check
→
Calibration4 cycles, signed exit
→
Release + self-exportreconstructable
06

Executive dashboards and content engine

Live, pilot ran
Problem
Status and decision information lived in documents that went stale between updates, so the executive view was always slightly behind.
Approach
Serve small web apps directly from workflow webhooks as branded HTML, backed by two databases that refresh automatically.
Result
Four live apps: a one-page CEO decision brief, an interactive status dashboard with inline editing, a posting calendar, and a status-update endpoint. Supported a CEO thought-leadership pilot end to end.
Build: no separate hosting, no separate deploy. Each app is HTML rendered by the workflow that owns the data, which means the page cannot drift from the source. The dashboard writes edits back through the same webhook. The content side shipped content pillars, a voice profile and a twelve-post calendar for the pilot.
Two databasesowned by the flow
→
Webhook renderbranded HTML
→
Brief, dashboard, calendar4 apps
→
Inline editwrites back
→
Auto refreshnever stale
🧠

Executive AI enablement

Onboarded the CEO, Chief of Staff and CTO onto an AI assistant: per-person context files built from corporate material, tone learned from their own writing, curated skill sets matched to each role, and a reusable setup guide for any new licence holder. Change management as much as engineering.

7 reusable agent skills, handed over
⏰

OKR reminder service

A scheduled workflow that reads the initiative board and emails each owner ahead of the twice-monthly update deadlines, weekend shifted, in the business time zone. Read-only by design: it never writes to the system of record. Pilot mode, forced runs and test mode live in a settings node, so a behaviour change is a flag flip, not a redeploy.

10 nodes, live to real owners

Built for my own ventures

The same engineering, on systems I own outright. These are where I test a pattern before it goes near a client.

🎬

MeisterSmith content engine

My own e-commerce content venture, running as one pipeline: product research and scoring, script generation, text to speech, AI b-roll, karaoke captions, music, and a scheduled multi-platform publish. Deterministic rendering with model-written copy, the same split I use on client work.

Own venture, runs daily
🧾

Email to invoice pipeline

Invoices arrive by email as PDFs. OCR extraction, an AI validation pass, an exception branch for anything the model is not confident about, then automated filing to cloud storage with a ledger entry. Humans only see the exceptions.

Estimated 60% less manual handling
💬

LinkedIn content engine

A multi-agent pipeline that gathers topics, drafts posts, scores them against quality gates, and publishes on a schedule with engagement capped and tracked. Reasoning sits in the agents, execution sits in Python, and a review gate stands between the two.

Own venture, scheduled daily
📱

Telegram operations briefings

A notification service that pushes a morning calendar brief, hourly reminders and content ideas to Telegram on a schedule. The day opens with the state of things instead of a search for it.

Own venture, running

Want one of these pointed at your process? Bring me the one that eats the most hours and I will tell you what it costs to remove.

Book a scoping call
Engineering practice

Anyone can wire a prompt to a webhook. The difference is what happens when it fails.

These are the rules I build to. They are the reason the systems above are still running.

📜

Nothing breaks for long

Every system I ship is version controlled with automatic backups and rollback checkpoints. If an update goes wrong, I roll back in minutes. Your operations do not stop while something gets fixed.

📏

The AI never holds your numbers

Anything factual or numeric in a report comes straight from your data. The AI only writes the words around it. That removes the classic AI failure of invented figures reaching a document your clients see.

🔒

No invented links or facts

Anything the AI produces that points to the outside world is checked against an approved list before it can reach output, and checked again at the final step. An invented link simply cannot get through.

🛡

Tested by attacking it on purpose

Systems that read outside input, like emails and web forms, get attacked deliberately before launch, so a malicious message cannot hijack the automation. Roughly 38 attack and edge-case scenarios run across two systems, with zero leaks.

⚖

Bad days degrade gracefully

When a third-party service goes down, the system delivers a thinner result and tells you exactly what is missing, instead of failing silently or stopping your operations. Failures are loud, small, and recoverable.

💰

Your AI bill stays predictable

Caching and model selection are built into every system, so repeat work costs nothing and the cheapest model that can do the job correctly is the one that runs it. Running cost is a design decision made up front, not a monthly surprise.

The part most people skip

Governed delivery

A proposal comes before the build: problem, value, architecture, MoSCoW scope, risks, phased delivery, and numbered open questions you can answer one by one. Then weekly board review, UAT on your data, and probationary monitoring after go-live. Credentials belong to a project, never to a person. Access is read-only unless write is genuinely required. A human reviews anything that reaches an external audience. This is the difference between an automation that survives its first bad week and one that quietly gets switched off.

Stack

What I actually build on, across client engagements and my own systems.

AI models and APIs

Claude APIGPT / OpenAI APIGemini OpenRouterGroqWhisper Google Flow / Veo

Orchestration

n8n (self-hosted)Claude Code agents and skillsMulti-agent pipelines PlaywrightPythonJavaScript PowerShell

Business integrations

Microsoft Graph / M365HubSpot CRMApollo.io Jira / Confluence / JSMNotion APIGoHighLevel MailgunTelegram Bot API

Media and documents

FFmpegMoviePyPIL / Pillow reportlabopenpyxlpython-pptx

Delivery and infrastructure

Webhooks + HTML appsCloudflare Pageslaunchd scheduling gitAutomated test suites

Engagement practice

SpecificationsMoSCoW scopeAdversarial testing UATRunbooksRollback checkpoints
Pricing

Three ways to start, with the number on the table

Fixed price, agreed before the build begins. Scoping is free, so the first conversation costs you nothing.

Essentials

$1,299
About 20 hours of build

One workflow, automated end to end. The single process that eats the most time, removed first so you can see what the rest is worth.

  • Scope: one workflow, automated end to end
  • Delivery: about 1 week
  • Payment: 50% at kickoff, 50% at handover
  • Included: user testing, documentation, two weeks of support
Most common

Professional

$1,999
About 40 hours of build

A multi-workflow system with 2 to 3 tools integrated, so work moves between your systems without a person in the middle keeping them in sync.

  • Scope: multi-workflow system, 2 to 3 tools integrated
  • Delivery: 2 to 3 weeks
  • Payment: 50 / 25 / 25 on milestones
  • Included: user testing, documentation, two weeks of support

Enterprise

$2,799+
About 60 hours of build

An end-to-end operations system, documented, with your team trained to run it. Built to survive the day I am no longer in the conversation.

  • Scope: end-to-end ops system, documentation, team training
  • Delivery: about 4 weeks
  • Payment: 50 / 25 / 25 on milestones
  • Included: user testing, documentation, two weeks of support

Time-based consulting

Let's talk

One-time consulting: I review your systems, troubleshoot, and hand you recommendations and a written report. Longer committed engagements, up to full time, are also on the table. Defined-scope builds are always quoted fixed price. Contact me and we will shape the engagement and the rate around your needs.

  • Scoping is always free.
  • Larger scopes are split into phases, and each phase is priced as its own project.
  • Every build includes user testing and a two-week post-handover support window.

Pricing as of August 2026

Not sure which tier fits? That is exactly what the scoping call is for, and it costs you nothing to find out.

Book a scoping call
FAQ

The questions that come up before every engagement

How does pricing work?

Fixed price, agreed before the build starts. I scope the work on a free call, then quote against one of the three tiers on this page, so you know the number before anything is built. There is no meter running and no invoice you did not see coming.

How long does a build take?

One to four weeks, depending on the tier. Essentials runs about a week, Professional two to three weeks, and Enterprise about four. Anything larger is split into phases and each phase runs on that same clock.

What happens after handover?

You get the workflow source, written documentation and a walkthrough video with whoever will own it. Two weeks of post-handover support is included in the price, so the first real week of running it is covered. A monthly retainer after that is optional and only makes sense if you want ongoing changes.

Do you work with agencies or subcontract?

Yes. I hold ongoing agency engagements and I am comfortable building under someone else's brand and delivery process. I plug into your existing intake, review and reporting rhythm rather than asking you to adopt mine.

What tools do you build on?

Self-hosted n8n for orchestration, LLM APIs from Anthropic, OpenAI and Google for the reasoning, and Python for anything that has to be deterministic. Integrations run through Microsoft Graph, CRM and ticketing APIs, and email infrastructure. The full stack is listed on this page.

Can you show client work?

Client engagements are covered by confidentiality, so every system is described by what it does, how it is built and what it changed, never by client name or client data. The case studies on this page are written that way on purpose. On a call I can walk through architecture and design decisions in as much depth as you want.

Documents

Everything you need to decide

Two PDFs. The deck if you want the overview, the resume if you want the background. Pricing is on this page, just above.

📊

Capabilities deck

The client-facing overview: what I build, how an engagement runs, the systems already shipped, and what happens at handover.

Open PDF →
📄

Resume

Full background, engagement history and stack, in the standard format if your process needs one.

Open PDF →

Let's scope it

Bring me one process that eats hours every week. I will tell you what it costs to remove, and whether it is worth removing. Scoping is free.

Book a scoping call Pricing Capabilities Deck LinkedIn