AI Operations

Meindert Smith

Digital Employees · In Production

I build digital employees. Not chatbots. They recall everything your company has ever produced, then do the real daily work: pricing, estimates, invoices, bookings, email. You hold the approvals; I keep it running. Scroll. The story is behind this page.

scroll
Stage 01 · Weeks 1–4 READ EVERYTHING

First, the chaos.

This drift is your company today: every email, file, invoice, image and calendar entry. Hundreds of thousands of artifacts, findable only by whoever remembers where they are. Watch them stream into one place, on dedicated hardware. Not a hyperscaler, not a data broker.

Stage 02 · Weeks 3–8 MAKE IT ASKABLE

Then, a memory.

The artifacts organize: deduplicated, text-extracted, classified, and linked into one connected corpus. Structured facts plus meaning-vectors, every answer traceable to its source. Ambiguity queues for a human instead of guessing.

Stage 03 · From week 6 GIVE IT A NAME

Then, an employee.

The memory becomes a colleague with a name. These two are Elliot and Sunny, answering their teams today. Citations on every claim, write-access earned one guarded verb at a time.

Stage 04 · Ongoing RUN IT LIKE PAYROLL

Then it works, every day.

Weekly capability shipping, a nightly canary that proves the critical safety path still fires, encrypted backups proven by actually restoring them, and a living handbook so the owner always knows exactly what their employee can do. Meet them below.

01

Two employees, on the job today

Each is a living system with its own name, memory and responsibilities, and its own always-current operations hub the client's whole team reads. Interactive tours are coming online here.

02

What I built, and how it's proven

The difference between a demo and a digital employee is what happens when things go wrong. These are the protections that keep the system working after launch: each one a failure I hit or saw coming, the mechanism that now prevents it, and the check that tells me it's still alive. Not promises. Promises don't survive contact with production.

The model proposes, code disposes
Agents don't free-form their output. They call bounded Python verbs that do the actual work, so the same request produces the same result whether the model was sharp that day or not. Choosing the verb is the model's job; everything downstream of that is deterministic.
Tiered review: independent, not sequential
High-stakes extractions get reviewed a second time, blind to the first verdict, so the reviewer can't anchor on it. Disagreements escalate to a third opinion as tie-breaker instead of being averaged away.
A model swap has to earn its place
Model changes are gated behind a scored harness run on the real corpus, not vendor benchmarks. The last challenger lost to the incumbent and was rejected. Regressions get caught before they ship, not after someone notices bad output.
Multi-modal extraction, routed by content
Documents are routed by density: real text layers are parsed directly, scans and photographs go to a vision pass that reads classification and fields in one shot. Reader health is monitored, so a silently failing extractor surfaces instead of quietly emitting nothing.
Guarantees are tested like features
Safety paths get the same treatment as product code: the block logic carries a regression suite, and a nightly canary drives a real violation through the live system and alerts if it survives. That's how I found a gate that reported healthy and had never once fired. A protection nobody exercises is indistinguishable from one that's broken.
The system catches what I forget
Vendor updates silently overwrite local patches, so watchdogs re-apply them nightly and alert on drift. A coverage lint flags any patch I forgot to register for that protection. The failure mode it guards against is me.
03

Your business could hire one

If your company's knowledge lives in inboxes and heads, that's the problem I solve. Tell me about your shop.

meiniesmith@mac.com →