Home / About
About
Why trust this record: how I think, how I got here, and why I publish the work in the open.
I'm Sean Duncombe. I design and build custom AI workflows and the internal systems businesses quietly run on — scheduling, automation, and Google Workspace integrations — through my practice, Neuroframe.
The work is mostly invisible when it goes well: a calendar that never double-books, an intake process that routes itself, a permission model a security team signs off without a second meeting. What's worth judging isn't the feature — it's the reasoning underneath, and whether it survived production. So this site is the record of exactly that: every system with its decisions, incidents and revisions, written to be audited rather than admired.
How I think
Decisions before code. I write the RFD and argue the trade-off before building. The reasoning outlives the implementation.
The boring mechanism wins. Between a clever fix and a dull, provably-correct one, I ship the dull one — it's what still makes sense at 3 a.m.
Make failure recoverable, then publish it. Not everything can be prevented; everything can fail safely and get an honest post-mortem.
Don't duplicate the platform's state. A second source of truth is a second thing to be wrong. I remove parts before adding them.
Done means someone else can run it. Documentation isn't the paperwork after the work — it's the proof the work is finished.
How I got here
Wrote the first Request for Discussion. Every system since begins as a numbered, argued RFD before any code.
Least-privilege domain-wide delegation across Calendar, Gmail and Drive — scoped to pass a client's security review rather than skirt it.
An event-driven job runner with per-run audit trails; recurring human processes became observable and retryable.
Scheduling on Google Workspace with no database of record. RFD-014 replaced compensation with per-practitioner serialization — the double-book stopped happening.
A daylight-saving incident drove a v3 rebuild and a DST regression suite; the same year, an AI-assisted intake layer went into production.
Began publishing the systems, decisions and incidents as this operating record.
What I'm working on now
There is usually something mid-way from private to public: extending the decision log into open view where client confidentiality allows, and hardening the systems already documented here. In time the contact flow on this site will run on the same booking engine described in its dossier — the record dogfooding itself.