Nox

AdeptOps · Long Beach, California

AI built around how your business actually works.

Reading the paperwork. Answering the questions your team keeps answering. Spotting the transaction that looks wrong. Keeping the numbers straight. Custom systems for the work that eats your week — built to your rules, and you can check every call they make.

80/80Formulas verified
0.0003%Maximum error
22/22Zero false rejects
01How the work actually works

Something arrives. A decision comes out.

If your business has a person opening files and typing what they say into another system, this is the loop that replaces the typing but keeps the judgement. Pick what lands in your inbox.

Same four stages every time. What changes between businesses is the fields and the rules, and those are yours.

Documents are where most people start, because that is where the hours are. The same intelligence does other work too: answering questions from your own documentation, scoring transactions for fraud, reconstructing rules nobody wrote down, and watching your numbers for the thing that just broke.

02The one I run myself

My agency runs on one of these.

Nox is the operations brain for AdeptOps. Not a demo and not a side project — it is the system of record this business actually runs on, and the same discipline I sell is the discipline it enforces on me.

CommercialClients, contacts, contracts, milestones, invoices and payments. The money side lives here, not in a spreadsheet.
DeliveryProjects, workflows, workflow steps and tasks. Every piece of client work is a row something can act on.
AgentsRegistered agents claim tasks and log runs against a shared event blackboard. Several can work at once without colliding.
MemoryLessons are embedded and searchable, so a problem solved once does not get solved from scratch again.
01
Agents never touch the database.

Every action an agent takes is a typed, logged tool call. There is no path from a model to raw SQL. If an agent does something surprising, there is a record of exactly which tool it called and with what.

02
There is a review queue.

Work that needs a person waits for a person. The same human-approval gate I build for clients is the one standing between an agent and my own invoices.

03
It is separate from every client system.

Its own database, its own vector store, its own container. It cannot reach a client's data, and a client's system cannot reach it.

Why this matters to you

Most people selling AI automation have never had to live with one. If a system I built started inventing invoices or losing tasks, it would be my business that broke.

So the guardrails on this site are not a philosophy. They are what I needed in order to trust the thing running my own company, and they are what you get.

Start with a diagnostic
03Verified results

Not claims. Proof.

Every number here traces to a specific piece of work. Nothing is rounded up, and where a figure comes from a test rather than production, it says so.

80/80 Formulas verified Across two independent schools, against official government rules.
0.0003% Maximum error Sub-cent rounding. Nothing else.
22/22 Zero false rejects On the regression set, across four extraction paths.
Case studies
Payroll logic extractionGRADOS · Argentina Reconstructed undocumented salary rules from PDF receipts alone, then proved the reconstruction against known answers. 80 of 80 formula matches. Read Support agentGRADOS · Argentina A production agent serving an organisation supporting around 82 schools, on WhatsApp and web. Escalates rather than guesses, and reviews its own work. Read Multi-tenant ad dashboardBakeano · Argentina Twelve client accounts on one platform, ingesting exports where no two files agree. AI insights cannot reach a client until a human approves them. Read Fraud detection & complianceCompleteSMS Applicant vetting and 10DLC compliance for a B2B SMS provider. 22 of 22 matched on the regression set, zero false rejects. Read
04What I build

Four ways to put AI to work.

01

Document Intelligence

The core. Extraction, classification, validation and rule inference from PDFs, spreadsheets and scanned paperwork. When the rules aren't documented, I reconstruct them from the outputs and prove the reconstruction is right.

02

Knowledge Agents

Support and intake agents that answer from your actual documentation, cite what they used, and hand off to a person when they should. Web, WhatsApp, or inside the tools you already run.

03

Operational Dashboards

Multi-tenant reporting that ingests the exports you already have, watches for anomalies, and tells you when something breaks before a client notices.

04

Detection & Review

Models that learn what normal looks like in your data and flag what does not: fraud signals, duplicate claims, accounts going quiet, numbers that stopped reconciling. Every flag arrives with the evidence behind it and a person decides.

05Method

Smart enough to read it. Honest enough to show you why.

The AI is doing genuinely hard work here. It opens a document nobody described to it, works out what each value means when every sender labels things differently, and infers rules that were never written down. That capability is new, and it is the reason these systems can do what automation software never could. What I build around it is the accountability.

  • The AI does the reading and the reasoning.

    Extraction, classification, inferring the rule from examples, finding the answer in a shelf of documentation, spotting the pattern that shifted. None of that is possible with rules alone.

  • Deterministic logic holds the final verdict.

    So the intelligence is accountable rather than unaccountable. Same inputs, same outcome, and you can trace exactly which rule fired on which value.

  • The AI can escalate, never approve.

    A model can push a case toward human review. It cannot clear something the rules rejected.

  • New behaviour runs in shadow first.

    It logs what it would have done, so its judgment can be measured against reality before anyone gives it authority.

  • Changes are proven not to break what worked.

    Responses on unchanged paths are verified byte-identical before anything ships.

DOCUMENTS SIGNALS ADJUDICATOR cited evidence RULES ENGINE SHADOW logs only PASS REVIEW escalate only
The model never holds the verdict
06Continuous improvement

It gets better on a schedule.

Most automations are frozen the day they ship. I build them to review their own work. On a recurring cycle the system reads back everything it handled and asks four questions about itself.

Q1

Where was I unsure?

Every answer carries a confidence score. Low-confidence responses are collected, not buried.

Q2

What had no answer?

Questions where nothing useful was retrieved are flagged as coverage gaps. A different, more valuable signal than being wrong.

Q3

Where was I corrected?

When an operator edits a draft, the system flags it when the proportion of changed text crosses a threshold, or when a critical value changes.

Q4

What keeps coming up?

Recurring gaps are clustered by topic and source document, so the fix addresses a pattern rather than one bad answer.

Then it stops and waits.

Each cycle produces a dated report and drafts the knowledge it thinks would close the gaps it found. Those drafts land in a review queue. A person approves, edits or rejects each one. Nothing enters the knowledge base without a human saying yes.

The agent proposes. It does not edit itself. A system that silently rewrites its own knowledge is a system nobody can audit.

RunsEvery Sunday, 08:00
Weekly reviews14
Proposals approved5
Period coveredMay – Aug 2026
Systems running it1 client system
Applied without review0
07Live agents

Talk to one.

Three agents I built, running right now. No signup, no sales call. Each shows a different piece of how these systems reason.

08Operations

Your servers or mine. Your choice.

Some clients want the whole thing hosted and run for them. Others want it inside their own network, on their own servers, with nothing leaving the building. I do both, and I will not push you toward either one.

Hosted with me, every client sits in their own isolated stack with their own database. Nothing is pooled, nothing is shared between clients, and nothing is used to train anything. Self-hosted on your side, the same system goes onto your infrastructure with the keys, the documentation and the runbook, and your data never leaves your control at all.

Tenant isolationdatabase level
Backupsnightly, rotated
Provisioningscripted
Monitoringhealth + outage alerts
09Reach

One standard.
Two languages.

I deliver in English and Spanish across US and Latin American markets. Not translated after the fact. Built bilingual, with interfaces, knowledge bases and reporting in the language users actually work in.

WhatsApp included. For much of Latin America, and for Spanish-speaking customers and workforces in the US, WhatsApp is where business happens. Shipped to production, integrated properly, with the same handoff rules as every other channel.

10How we work

Diagnose. Build. Maintain.

STEP 01

Diagnostic

A paid engagement ending in a written deliverable: which of your document workflows can be automated, what each costs you now, what it takes to fix, and what to do first. You own the report whether or not you hire me to build it.

STEP 02

Build

Scoped from the diagnostic. Delivered in phases with a working system at the end of each one, not a big-bang launch.

STEP 03

Maintain

Systems drift. APIs change, formats change, rules change. Ongoing operation, monitoring and improvement, on retainer.

Josh Dannhauser, founder of AdeptOps
11Who you'd work with

Josh Dannhauser

Founder, AdeptOps

I build AI systems for operations-heavy businesses. Before this I spent years in high-volume operations, which is where I learned that the expensive problem is almost never the interesting one. It's the four hundred PDFs.

I work directly with clients. There is no account manager and no junior team. The person who scopes your project builds it.

Based in Long Beach, California. Working across the US and Latin America.

12Get in touch

Tell me what's eating your time.

The more specific you are, the more useful my first reply will be. I read every one of these myself.

Response time1 business day
LanguagesEN / ES
BasedLong Beach, CA

Or email josh@adeptops.dev directly.