Agent Build
AI agents that run
real processes
One process automated, in production, with monitoring and error handling. Fixed scope. I build it myself, and I hand it over.
The problem
AI pilots rarely fail on the model
They fail because nobody owns the process underneath. Because there's no agreement on what happens when the agent is wrong. Because it was built by someone who didn't know the business process, or designed by someone who didn't know the technology.
I sit on both sides of that table. I've run operations at C-level and I write the code. That means the question "what do we do when it's uncertain" gets answered before the build starts, not after.
What I deliver runs in production or it isn't finished. An agent that only works in a demo isn't an agent, it's a presentation.
Proof
I build this for myself too
Sentinel is an always-on monitoring and synthesis agent on my own stack. It reads across Office 365, Close CRM, GitHub, Azure DevOps, PostHog, Notion and Billit, keeps a persistent ledger of open loops, and classifies every proposed action into three tiers: act, ask, or don't touch. Escalation runs over Telegram. Access sits behind Tailscale, because the data is sensitive.
Hermes is an internal AI service with a chat frontend for the team and a bidirectional CRM integration, with RAG over the company's own knowledge.
That isn't a side project to look impressive. It's the same architecture I apply with clients: read broad, write narrow, explicit guardrails, and a human in the loop on anything irreversible.
Where this works
Processes that lend themselves to it
Reconciliation and control
Billing against CRM, payments against accounting, contract against delivery. Dull, repetitive, error-prone, and exactly what an agent is good at.
Reporting and synthesis
Weekly overviews, board packs, status reports. Pulling data from five systems and turning it into a readable narrative.
Monitoring and follow-up
Catching signals across systems, connecting them, and escalating before anyone notices by themselves.
Back office and admin
Creating records, moving data, updating statuses, chasing missing information.
Onboarding and customer follow-up
Who is stuck, where, and for how long. Followed up automatically instead of discovered monthly.
Unlocking knowledge
RAG over your own documentation, contracts and history, so the team stops searching.
Approach
How an Agent Build runs
Lock the scope
One process, sharply bounded. We agree up front what the agent may decide, what it must escalate, and what it never touches.
Build against real data
No synthetic test set. I build against your actual systems, in an environment where mistakes break nothing.
Guardrails and observability
Logging, error handling, escalation paths and a clear threshold above which a human decides. This is the part pilots skip.
Rollout and handover
In production, with documentation and a team that knows how it works and how to change it. I'm not a dependency.
Stack
What I work with
Field guide
From AI pilot to production
Twelve questions to answer before you build, what “in production” actually means, and why most AI projects stall right there. Eleven pages, no sales pitch.
No email address required. Just download it.
Practical
What you probably want to know
What does an Agent Build cost?
A fixed price per process, agreed before the build starts. No open-ended time and materials. The scope is written down: what the agent does, what it escalates, and what it never does. Anything outside that is a separate proposal, not a change-request invoice.
How long does it take?
For one scoped process I count in weeks, not months. Exactly how long depends on how many systems hang off it and how clean the data is. I'll estimate that in the first call.
Do I have to send my data to an AI vendor?
Not necessarily. Where sensitivity demands it, it runs self-hosted, behind your own network. My own agents run that way. We decide that up front, not afterwards.
What if the agent gets it wrong?
That's not an edge case, it's a design question. Every agent gets a threshold above which a human decides, logging of what it did, and an escalation path. That part is exactly what pilots skip.
Will I depend on you afterwards?
No, and that's the point. Code, documentation and ownership go to your team. I stay available if you want, but it has to run without me.
Got one process in mind?
Tell me which process annoys you most. In thirty minutes I'll tell you whether it suits an agent and roughly what it takes.
I usually reply the same day.