Sorellan & Kairoth

Agentic systems that hold up under audit.

An engineering lab — run by people who have been the operator, the CFO, and the auditor — that designs, builds, and operates agentic systems for the offices that run real operations.

We are not a consultancy. We do not sell decks. We do not run pilots. We build the system, hand it over running, and operate it if you want us to.

↓ What we build

A representative audit record: the disbursements agent (version 4) auto-approved a payment because the vendor is on the whitelist, the invoice is within 2% tolerance, and the general-ledger code matches the prior fourteen invoices — confidence 0.96, approved by the controller under policy version 4, written to an immutable, timestamped record.

What we build

The operating layer, not another feature.

A tool answers a question. An operating layer runs the work — it reads the messy inputs, plans under your constraints, acts across the systems you already own, verifies the result, and writes a record you can defend. One queue, every domain, one audit trail underneath.

  1. Intent
  2. Plan
  3. Execute
  4. Verify
  5. Record
Who builds this

Most agentic systems are built by engineers who have never closed a month, signed an audit opinion, or owned the P&L. We have.

01 Auditor Big Four. Sat across the table; signed the opinion.
02 CFO · C-suite Owned the close, the controls, the board number.
03 Operator · Founder Ran the operating company, not a slide about one.
04 Engineer Now we build the systems we wished we'd had.

So the guardrails aren't theoretical. They're the controls we'd have demanded from the other side of the table.

How we build safely

Defensibility is the design — not bolted on after.

Operator

Bounded autonomy. Nothing consequential runs without a human; the team works the exceptions, not the routine.

We ran the operation

Auditor

Every action is a row — plain-English reasoning, every evidence source cited, a confidence score, the human who approved it. Immutable and timestamped. The record is the walkthrough.

We sat for the audit

Risk

Anomalies caught by context, not keywords. Per-entity behavioral baselines. The kill switch fails closed.

We owned the downside

Compliance

Policy-as-code, versioned. Every decision remembers the exact policy in force when it was made.

We signed the rep letter
Models & tooling

Built on the frontier. Run like infrastructure.

We match frontier models to the task the way you would staff senior and junior people, then wrap them in a Fortune-100-grade engineering stack. Credentials are scoped, rotated, and logged; agents hold no standing access. Your data stays inside your boundary.

  • Immutable
  • Timestamped
  • Replayable
Engagement

An engagement begins in writing.

There is no intake call. Correspondence is in writing, by introduction. Send the brief — the first reply is the work.