Work

Systems we run, described at architecture level.

These are pieces of Skeen's own practice. Each note covers the problem, the approach, the stack, and the outcome we can state without inventing numbers or naming clients.

  • Multi-Agent Governance

    Coordinator + Specialist Lanes · 100% Human Egress Gate

  • Zero-Trust Isolation

    Per-Origin Service Tokens · Signed Scope Files

  • Privacy by Architecture

    CSP connect-src 'none' · Verified Encrypted DR

  • Enterprise Data Core

    Oracle DBA & PL/SQL · Informatica/ODI ETL · Maximo EAM

Multi-agent orchestration & human approval gates

A 30+ specialist agent fleet routed by a Chief-of-Staff coordinator onto a deterministic review queue. A person approves every external action.

Problem
An unattended send is hard to undo: the wrong recipient, a commitment nobody read, a message that cannot be unsent. The delay of a review is smaller than that failure.
Approach
Specialists do bounded work—research, classification, and drafts—and hand off structured packs. Agents are partitioned by role and blast radius: untrusted inbound parsing, internal system-of-record reads, and external egress stay in separate lanes, so no single agent holds the lethal trifecta of untrusted input, privileged credentials, and an autonomous external send. A Chief-of-Staff coordinator routes that work onto a deterministic queue a person can see. Read-only monitors may run. External actions wait. The layers exist so a draft cannot become a send by itself.
Outcome
This is the operating shape. We do not publish volume or time-saved figures.

In practice

  • 30+ Specialist Agents
  • Coordinator Routing Queue
  • Blast-Radius Partitioning
  • Mandatory Human Egress Gate
  • Structured Handoff Packs

Text version: specialist agents do bounded work, a coordinator places it on a review queue, a person approves, and only then does an external action go out.

Per-origin Zero Trust access to systems of record

Agents query operational systems through narrow, per-application service tokens and read/write split lanes—never shared long-lived credentials.

Problem
A credential pasted into every client fails all at once. One copy leaks, and every copy is exposed. An agent that can also send makes that leak an action.
Approach
Each application gets its own Cloudflare Access service token, scoped to one origin, so a token for one system cannot open another. Monitoring lanes are read-only GET. Anything that writes or sends waits for a person. Tokens are not copied into agent prompts: a local loopback bridge (127.0.0.1) holds the credential at the edge. Expiry is watched on a 90-day clock, with an alert before a token lapses. There is no shared long-lived bearer secret. The public site is not that path, and the inventory of connected systems is not published here.
Outcome
Narrow, per-origin access with a human gate on mutation. The inventory is not on this site.

In practice

  • Per-App Service Token Isolation
  • Zero Trust Access Policies
  • Read-Only vs. Write-Gated Lanes
  • 90-Day Expiry Alerting

Text version: untrusted input is sanitized, a specialist agent works without raw secrets, a per-application Zero Trust service token opens one origin, read-only monitors stay on GET while writes wait for a person, and only then is a system of record touched.

Private document lookup

Agents can query a private corpus. The corpus stays on systems we operate, and a person decides what leaves.

Problem
Drafts and research are more useful when they can see the actual documents. Copying that corpus to a public model is a leak you cannot take back.
Approach
The documents stay on systems we operate, separate from this public site, and are not copied to a public model for training. Agents may query and draft against them. A person decides what is sent or filed onward. Separation is the control: the marketing host should not be where the corpus lives. Operational state and document metadata are stored in offsite snapshots encrypted with age. Recovery is marked green only after an automated ephemeral PostgreSQL restore checks table and row counts.
Outcome
Lookup is possible without handing the corpus to a public model. Document counts are not published.

In practice

  • Self-Hosted Document Corpus
  • Zero Public-Model Training Leak
  • Encrypted Offsite Backups (.age)
  • Automated Ephemeral Restore Verification

Text version: private documents stay on systems we operate. Specialist agents may query them and draft. A person decides what leaves. The corpus is not copied to a public model.

Opportunity intake pipeline

Public-sector procurement and RFP signals are collected, scored, deduped, and placed where a person already reviews work.

Problem
Notices arrive from many public places, in different shapes, often more than once. Reading them raw is where a duplicate gets chased and a real fit gets skipped.
Approach
Ingest from more than one public source, score fit with a deterministic rubric, dedupe, and import into the system of record a person already uses. A daily digest shows what changed. Anything that would leave the practice still waits for approval. We do not list the sources here.
Outcome
Review happens on a queue, not in a pile of tabs. Counts are not published.

In practice

  • Multi-Source Ingestion & Dedupe
  • Deterministic Fit Scoring
  • CRM Queue Integration
  • Daily Digest Briefing

SkeenSecure+ (Dealership-model posture briefing)

A dealership-model network and security posture briefing in the client’s own environment, where the client holds the keys. Ted Skeen, Tampa, FL.

Problem
A small office is handed a 300-item scanner dump, or a contract where an outside vendor keeps the credentials, and still cannot say which few fixes matter this month.
Approach
Map first under a cryptographically signed scope authorization file, verified before any discovery probe runs, then run read-only posture checks in client-owned accounts. Workstation walk-away lock hygiene covers unattended screens and Bluetooth, with a written risk-acceptance log for what is fixed, accepted, or scheduled. The client holds the keys.
Outcome
Offered by Ted Skeen as a posture briefing and an estate map. Never a penetration test, a compliance certification or audit, or 24/7 monitoring.

In practice

  • Cryptographically signed scope files
  • Client-owned environment
  • Read-only posture checks
  • Prioritized remediation and risk-acceptance notes

Visit Brief (Zero-network patient & caregiver brief)

A free bilingual (English / Español) offline web app that prints one US Letter page before an appointment. 100% on-device. Not medical advice.

Problem
A short appointment gets spent reconstructing medications, allergies, and questions. Consumer health apps often ask for an account and a cloud copy of those notes.
Approach
A static offline app. Notes live in IndexedDB on the device. An optional file export is passphrase-encrypted with AES-256-GCM. Labels are English and Spanish. Care-Team Mode keeps more than one profile on the same phone, and text can scale to 200 percent. The hosting policy sets connect-src to none, and a Playwright check fails if a request leaves the page origin.
Outcome
A deterministic one-page US Letter printout: current medications, boxed allergies, what changed since last visit, and ranked questions. Paperwork and printing only — never medical advice. Nothing typed is sent to a server.

In practice

  • TypeScript offline web app
  • IndexedDB and Web Crypto (AES-256-GCM)
  • Content-Security-Policy connect-src none
  • Playwright check for requests that leave the page

Oracle Utilities WAM, IBM Maximo, and PL/SQL

Deep, hands-on Oracle database, enterprise ETL, and asset management delivery. The engineer who designs the change can support it in production.

Problem
Asset-heavy operations depend on Oracle data and an EAM system of record. A change that only exists in a design note fails the first time something breaks and the author is gone.
Approach
Delivery is hands-on: Oracle lead application development and DBA practice (PL/SQL packages, procedures, tuning, migrations, backup and recovery, patching, monitoring, hardening); enterprise ETL and reporting pipelines (Informatica, OWB, ODI, OBIEE, shell + PL/SQL); Oracle Fusion Finance and SCM support and data conversion; and IBM Maximo / Oracle Utilities WAM (work orders, job plans, asset hierarchies, data quality, conversion, and cutover). Packages are written to be tested, including with utPLSQL. Notes ship with the change so the next engineer can operate it. Production multi-agent AI orchestration has run beside this delivery since mid-2025. Agents draft around the schema. A person owns the change.
Outcome
Client names, system identifiers, and measured results are not published here. Those details belong in a conversation, under the terms of the engagement.

In practice

  • Oracle Database & DBA
  • PL/SQL & utPLSQL
  • Informatica / OWB / ODI / OBIEE ETL
  • Oracle Fusion Finance & SCM
  • IBM Maximo & Oracle Utilities WAM
  • Data Conversion & Cutover

This site

A static site. Each change is reviewed before the public site updates.

Problem
Editing the live site in place means the mistake is already public. A copy step onto a server is easy to skip and hard to review.
Approach
Every change is built and previewed before it goes live. A person reviews it, and only then does the public site update.
Outcome
Production changes only after a preview and a review.

In practice

  • Reviewed changes
  • Preview before production