Expertise

AI Workflow Automation

Designing multi-step AI workflows that combine generation, validation, human review and safe actions.

Problems I solve: Repetitive content and support workflows, Multi-step AI processing, Human-in-the-loop approval, AI generation with verification, Connecting AI outputs to practical actions
Approach: Break workflows into explicit steps, Separate generation from evaluation, Add review, apply, reject and retry states, Record outputs and decisions, Prevent irreversible actions without confirmation
Strengths: Review/apply/reject content workflows already running in production (this CMS's own AI Studio), Scheduled/automated jobs with deterministic fallbacks when LLM calls aren't warranted

Full description

I design AI workflows that do more than generate text. The workflow may classify, retrieve, draft, evaluate, revise, request approval and trigger an action. I focus on keeping workflows understandable, reusable and safe.

Capability statement

I break AI-assisted workflows into explicit states with review/apply/reject steps, rather than one opaque prompt-to-output call.

Current focus

Refining how the content-review workflow scores and surfaces AI-generated drafts before publish.

Internal Links

Keep exploring

Explore with Alex OS

Alex builds AI workflows with explicit states — classify, draft, evaluate, approve, act — so each step is auditable and reversible before anything irreversible…

AI Workflow Automation

Most AI integrations collapse everything into one prompt-to-output call. Alex structures workflows as discrete states: generation, evaluation, review, apply/reject, and action. Each state is explicit, recorded, and interruptible. The result is a workflow you can inspect, retry, or override — not a black box.

How the workflow is structured

Running in production: this CMS's AI Studio

The review/apply/reject content workflow is live inside Alex's own CMS (Alex CMS). AI-generated drafts are scored and surfaced for human review before publish — the current focus is refining how that scoring prioritises which drafts need attention first. The AI workflow automation proof documents the generation pipeline with cost tracing attached. The AI content scoring experiment shows the scoring logic in isolation.

Why separating generation from evaluation matters

Combining generation and evaluation in one prompt means a failure in either is invisible. Splitting them lets you swap the evaluator, adjust thresholds, or skip generation entirely when a deterministic fallback is cheaper and more reliable. Alex's scheduled jobs include explicit fallback paths for cases where an LLM call isn't warranted — the workflow decides, not the prompt.

Where this connects across the stack

Alex OS (the FastAPI reasoning and retrieval layer) sits between the CMS and the public site, handling grounded answers and journey-aware responses — a live example of AI outputs wired to practical actions with retrieval and context management built in, not bolted on.