AI at Basëthic
AI does the work. Rules decide what is allowed.
Most AI projects fail the same way. The model drifts from scope, invents a record, or sends something nobody approved. I build the control layer first, then put AI inside it.
Schedule a Systems Intro CallThe problem
AI is probabilistic. Your books, your CRM, and your customer emails are not. The gap between the two is where money and trust get lost. Prompting harder does not close it. Architecture does.
How I control it
Every AI-assisted workflow I build follows five rules.
1. A person approves anything customer-facing
Drafts are drafts. Nothing reaches a customer, an employee, or a financial system until someone signs off. This is written into my Terms of Service.
2. Reads and writes are separate
AI can read freely. Anything that changes a record goes through a defined step.
3. Every write is verified
After a record changes, the system reads it back and confirms it matches.
4. Math is done by scripts
Totals, hours, and reconciliations are calculated and checked in code, not estimated by a model.
5. Records are matched before they are created
Phone, address, and name are checked first, so AI does not create duplicates.
DatumLock
DatumLock is the control layer behind how I build. It sits in front of AI-driven work: a TypeScript router that accepts eight whitelisted actions, moves every change through Draft, Review, Approve, and Commit, and treats committed records as immutable. Changing one requires a formal revision. 670 automated tests cover it, and the router runs live on my own server.
Mergestalt
Mergestalt is the general-purpose operating framework I am building. CRM, scheduling, invoicing, and correspondence run on self-hosted Baserow and n8n, with Claude classifying incoming email, drafting routine messages for review, and turning a supervisor’s notes into formatted crew instructions. Client data stays on infrastructure I control, in a workspace isolated to that client. The first deployment is at a landscaping company.
LociOS
LociOS is the next system on the roadmap. It is a knowledge system built on the Method of Loci. Ideas are captured as thoughts, shaped into structured artifacts, connected by typed relationships, and organized into conceptual locations, so related knowledge lives in one place. The goal is resurfacing: bringing prior reasoning back when it becomes relevant instead of waiting for someone to search for it. It preserves where a decision came from, not just what was decided. The architecture is designed, and the build follows DatumLock. It is separate from both DatumLock and Mergestalt.
Where each piece stands
Running
- Mergestalt: lead intake with contact matching before creating, on self-hosted Baserow
- Mergestalt: QuickBooks Online sync, with customers created from leads and approved estimates turned into jobs
- Mergestalt: Claude summaries of estimate dictation and photos, saved to the CRM
- Mergestalt: daily briefing page combining the job schedule, calendars, and QuickBooks
- DatumLock: control router, live on my own server, called by an n8n workflow that has no decision logic of its own
- DatumLock: Notion and ClickUp adapters, covered by automated tests
- DatumLock: read-back verification on every external write
- DatumLock: drift detection, atomic rollback of failed commits, and execution boundaries enforced in code
In build
- Mergestalt: Claude email triage and correspondence logging
- Mergestalt: Connecteam scheduling sync
- Mergestalt: crew instructions, invoice preparation, and timesheet reconciliation workflows, handled by hand today
- DatumLock: chat interface
Planned
- LociOS: a knowledge system built on the Method of Loci that organizes thoughts into structured artifacts and resurfaces prior reasoning when it becomes relevant
Cost is a design decision
High-volume steps like email classification run on the smallest capable model. Larger models are reserved for steps that need them and never run inside per-record loops. On a 3,000-message backfill, that choice is roughly $4.50 versus $22. Multiplied across a year of operations, model choice is a budget line.
Where this fits your stack
AI works best added to the systems you already run, not bolted on beside them. Whether your team uses AI informally or has not started, I can put it inside the same controls: approval before anything customer-facing, verified writes, and a record of what AI did and who approved it.
What that looks like in practice:
- Classify and route inbound email, calls, and form submissions into your CRM, helpdesk, or work queue
- Draft replies and follow-ups for human review, never auto-sent
- Extract data from invoices, receipts, worksheets, and other documents into your CRM, accounting system, or database
- Turn notes and voice memos into formatted instructions, summaries, or records
- Reconcile records across systems, with totals and matches verified by script
- Summarize account or job activity into the channel your team already reads
HubSpot, QuickBooks, Baserow, Notion, n8n, or a custom API: if it has an API, there is usually a path.