BridgeSEO | Frederico Luz - AI Engineer & Agent Architect

Completed article view showing word count against target, hero image direction, and final publishing checks marked published — A finished article carries its evidence with it: word count against target, hero-image direction, and final publishing checks. Client attribution in the byline was removed for publication here.

More product screens (2)

BridgeSEO content writer workspace showing a structured article brief with topic, keyword, audience, goal, and editorial angle fields — The writer workspace shapes a structured brief — topic, keyword, audience, goal, and must-cover points — before generation. Product UI shown under its GrowthOS branding.

BridgeSEO audit setup with site health, content review, growth opportunity, AI visibility, and custom audit options — Choose the scope before running an audit: site health, growth opportunities, content review, AI visibility, or a custom audit.

Problem

Hexa Intelligence needed to produce SEO content for multiple clients without treating an LLM's first draft as a finished article. BridgeSEO is the answer: a staged pipeline that validates the brief and client context, gathers knowledge and does SEO research, plans the article, writes it in bounded batches, generates schema metadata, audits the result, remediates targeted failures, and syncs approved output to WordPress.

Each phase is independently routable, so research, writing, planning, auditing, and remediation can run on different models instead of forcing one expensive model onto every task. Supabase is the source of truth; WordPress is the publishing layer.

Pipeline diagram: a validated brief with client context moves through research, planning, and writing in bounded batches with schema metadata, then a bounded audit-and-repair loop across four audit dimensions, editorial review, and sync to WordPress. Supabase holds article state, run history, and per-phase usage; each phase can route to its own model. — Each phase is a separate, routable step; audit and repair loop a bounded number of times before editors review.

Role and ownership

As a freelance Product Engineer / AI Engineer I owned architecture, implementation, model and pipeline choices, evaluation instrumentation, reliability fixes, deployment, and technical communication with Hexa leadership — turning broad requirements into bounded pipeline phases and reviewable product behavior.

Engineering decisions

Compose the pipeline from inspectable phases. Brief validation, research, planning, section writing, audit, remediation, schema, and publishing each carry their own model routing, progress, and telemetry rather than one opaque completion.
Audit across four dimensions. Content/E-E-A-T, GEO/AI visibility, on-page SEO, and schema are scored separately so a single opaque number cannot hide a structural or editorial failure.
Make repair safe. Word-count floors, empty-section checks, structured-output hardening, phase-aware timeout budgets, bounded retries, content-loss protection, and actual-served-model attribution keep a malformed completion or silent provider fallback from being marked successful.
Treat model choice as an evidence problem. Per-phase usage events record model, tokens, duration, and cost, so routing and billing decisions are testable rather than assumed from a headline price.

Evaluation and result

An evaluation group ran the same brief through multiple model configurations and displayed comparable outputs with quality scores, word count, latency, tokens, and cost. That like-for-like comparison reduced measured per-article generation cost from $0.73 to $0.09 with no human-rated quality loss — a result of instrumenting the workflow and comparing matched runs, not of swapping a model and assuming quality held. The figure is scoped to that writer-model comparison, not a claim about the whole pipeline's price.

Evaluation diagram: the same brief runs through several model configurations, and their outputs are compared on quality scores, word count, latency, tokens, and cost. In the writer-model comparison on matched briefs, measured cost per article fell from $0.73 to $0.09 with no human-rated quality loss. — Matched-brief comparison of writer models. The $0.73 → $0.09 figure is scoped to that comparison.

The same telemetry answered practical questions in production: which phase spent money, which model actually served the request after fallback, and whether a cheaper run still produced a usable article.

Status and boundaries

BridgeSEO is freelance platform work for Hexa Intelligence, ongoing since February 2026. It is distinct from my earlier Hexa internship — which built a health-content portal and the Carglass migration — and from the separate Santander Content Accuracy Monitor delivered through the same company. The platform later extended into social planning and image workflows; those are described as implemented work exercised on staging, not blanket production claims. The system prepares and routes work for editorial review rather than replacing it.