Marcos Souza

Maintaining user documentation and onboarding tutorials manually (taking screenshots, recording steps, and updating help articles) consumed dozens of engineering and product hours after every release, creating stale guides and heavy support ticket loads.

Created a custom AI agent skill that reads canonical Markdown user-journey scripts as the single source of truth. The AI agent executes the entire loop: boots a local web instance in development, autonomously navigates through the exact user flow via browser automation, captures step-by-step screenshots and videos at critical interaction points, and compiles the official 22-tutorial interactive Help Center.

Invested upfront in deterministic script definitions and browser automation tooling in exchange for near-zero ongoing documentation maintenance.

Autism (ABA) clinic therapists spent 1+ hours daily transcribing voice notes and formatting clinical session evolution reports, causing clinician burnout. Unconstrained LLMs risked dangerous medical hallucinations in clinical charts.

Asynchronous FastAPI microservices combining Whisper audio transcription with Claude reasoning. Enforced strict Pydantic output schemas matching ABA clinical taxonomy, paired with an intuitive Human-in-the-Loop review queue where clinicians review side-by-side diffs and explicitly approve records before database persistence.

Sacrificed fully automated 1-click publishing in exchange for 100% clinician auditability and zero liability risk in regulated medical records.

Manual QA before releases created 1-hour bottlenecks, slowing deployment cadence. Furthermore, using AI coding assistants without rigid verification risked silent business regressions ('vibe coding').

Engineered an automated E2E test harness using Playwright with ephemeral Docker Compose database fixtures, pre-seeded role accounts, and strict Page Object Models. Zero retries allowed in CI to enforce zero tolerance for flaky tests, creating a deterministic containment net for human and AI-assisted implementation.

Maintained strict fixture isolation in exchange for an unshakeable safety net enabling autonomous agents to safely propose verified PRs.

Routing multi-megabyte therapy audio/video recordings through web application API servers caused severe server memory exhaustion, high bandwidth egress costs, and 504 timeouts on slow clinic cellular connections.

Designed an edge pre-signed direct upload pipeline with Cloudflare R2. Locked maximum file size constraints (20MB) and cryptographic integrity checks directly into pre-signed URL extensionHeaders, completely bypassing application server memory during media transit.

Added client-side direct upload orchestration complexity to eliminate backend memory exhaustion and achieve zero egress costs.

During rolling zero-downtime production deployments, active users with open browser tabs triggered Server Actions pointing to stale build hashes, resulting in fatal 'Failed to find Server Action' errors and lost form state.

Implemented a client-side error boundary interceptor that detects deployment skew signatures, serializes dirty user inputs into temporary storage, and triggers a seamless state-preserving page reload. Middleware suppresses internal Next.js control-flow exceptions in Sentry.

Intercepted framework-level routing errors at runtime rather than forcing hard maintenance windows or user session cutoffs.

In a multi-tenant healthcare SaaS, an Insecure Direct Object Reference (IDOR) could expose confidential patient evolutions between competing clinics or unauthorized staff.

Layered authorization barriers at the service/BFF level with explicit organization ownership guards. Built visual permission matrix UI on staff forms, prevented duplicate organization memberships, and enforced strict therapist-to-patient assignment checks.

Enforced defense-in-depth permission verification on every query rather than relying solely on frontend route protection.