ET Ethos Tech / ethos-tech.ai

Work · show me, don't tell me

Systems we've built, and the patterns behind them

Real production systems and the reference architectures they're built on — described in shape, not logos. This is the discipline of AI use-case development, shown rather than claimed.

Systems

Quality architecture

Preflight

In production

The problem

LLM output reads fluent but is often generic, hedged, or wrong enough to burn credibility when it reaches a customer. A single shipped 'slop' email costs more than not using AI at all.

How it works

  • Layer 3 — a deterministic lint pass catches AI tells with regex: fast, explainable, repeatable.
  • Layer 2 — authenticity scoring, earned once the lint layer is catching real volume.
  • Layer 1 — generator / critic / editor isolation behind a hard context wall.
  • Phased on purpose: a persona only earns the next layer when its output volume justifies it.
Shipped plugin

Anti-Slop Linter

Shipped · installable

The problem

The cheapest, highest-frequency quality gate — catching AI writing tells deterministically — needed to be reusable by a whole team, not trapped in one person’s head or one chat session.

How it works

  • A packaged plugin: a deterministic lint pass plus a writing-style capture that learns a person’s voice rules.
  • Regex and string matching, so the gate is fast, explainable, and repeatable rather than another model call to second-guess.
  • Portable across Claude Code and Cowork with no rewrite.
Agentic OS

Odin

In production

The problem

A senior operator running work across many systems can’t reliably hold priority, follow through, or catch context shifts in time. On-demand assistants fail because replying is the first thing to drop under load.

How it works

  • Scheduled agents on their own cadence: a morning briefing, a retro that grades the day, a staleness detector, a weekly refresh.
  • Every surfaced item carries evidence re-derived from the live source before it reaches the operator.
  • It diagnosed its own failure mode — "confirmation decay" — from its tracker data, and the loop was redesigned in response.
Delegation design

Agentic Crew

Design-first · phased

The problem

One operator is the router for every piece of work — which capability to pick up, when, in what voice. That routing is the cognitive load; the goal is to take it off the plate without losing judgment or quality.

How it works

  • A phased portfolio of named personas, each a job with a contract: inputs, outputs, the decisions it can make alone, and the ones that escalate.
  • Map the work before you name a persona — no persona that can’t point to three real tasks from last month.
  • Four shared commitments every persona inherits: layered memory, a privacy-preserving bridge, phased anti-slop, and a shared verification primitive.

Patterns & methods

Reference architecture

Human-gated agentic pipeline

Teardown

The problem

A recurring, high-volume, customer-facing task where you want AI leverage but cannot let an unsupervised model touch a customer.

How it works

  • Score → Draft → Gate → Verify → Record. Agents propose; humans own every decision that reaches a customer.
  • A context-isolated checker re-derives the evidence behind every claim before it renders.
  • An append-only record turns every mistake caught at a gate into next round’s correction.
Read the full teardown →
Convergence

Loop engineering

Adopting portfolio-wide

The problem

Quality gates that run once and hand a draft back to a human still depend on the human to drive the next iteration. That manual back-and-forth is exactly the churn worth automating.

How it works

  • Convert one-shot gates into convergence loops: draft, run a fresh isolated checker, feed failures back, re-draft.
  • Stop only on a verifiable condition, with three HALT guards: repeat-finding, non-convergence, and an iteration cap.
  • Loop the iteration; keep the human on judgment. The gate sits at the end, never the middle.
Method

The 5S Framework

v2.9 · self-updating

The problem

Teams treat getting value from AI as a prompting problem. The hard part is everything around the prompt — and there was no repeatable way to design for it.

How it works

  • Scope — the outcome and whether AI is even the right tool.
  • Split — decompose the work into delegable pieces.
  • Shape — engineer the context the model runs inside.
  • Ship — deploy with gates and verification.
  • Sense — observe in production and log what to change next.

Let's talk

Have a problem shaped like one of these?

No forms. No qualification call. Pick a time and we'll talk about your actual problem.

Book the builder →

Prefer email? jeremy.brazell@ethos-ai.tech