AI automation in production: pipelines, flows and guardrails

In one line: Production AI here started with a translation pipeline in July 2024; durable Inngest workflows on tur-automations (Next.js on Vercel) now run scheduled digests, monitors, and site events — with Telegram approval before high-impact actions ship.

Context

Useful AI on this site and related properties is not a single chat widget. It is pipelines and orchestration: collect, transform, draft, validate, publish — with clear handoffs and retries when steps fail.

Why it was worth building

Manual translation and publishing burned calendar time and introduced inconsistency. Event-driven site flows needed a backend that could retry safely and stay observable without babysitting cron scripts on a laptop.

What I built

Starting point (2024)

• Translation pipeline (July 2024) — GPT-4o mini via OpenAI API for a multilingual site; Python for extraction, URL mapping, and validation; human review for tone and risk. Write-up.
• Slack-to-site publishing — structured Slack message → n8n → GitHub → deploy in under 30 seconds for this site. Case study.

Inngest on tur-automations (Next.js on Vercel)

Inngest is the orchestration layer: each workflow is a durable function with steps that can retry independently, triggered by cron schedules or events. That matters when external APIs, LLM calls, or deploy hooks fail mid-run — the platform resumes from the failed step instead of losing the whole job.

Production flows, in priority order:

1. Daily AI vendor digest — scheduled job that collects and summarises AI vendor and model updates worth reading.
2. Daily signal monitor — scheduled job that watches a configured set of sources and surfaces changes that may need a human follow-up.
3. Human-in-the-loop approvals via Telegram — automations and agents can propose actions (publish, merge, infra-impacting changes); a person approves with one tap before anything ships. Same guardrail idea as The Unnamed Roads agentic platform: autonomy with explicit gates, not silent production writes.
4. Daily synthetic site checks — scheduled monitoring of public sites for regressions that should not wait for a user report.
5. Event-driven site flows — newsletter signup handling, plain newsletter send, and contact-form notification when visitors submit the form on this portfolio.

Visual, branchy automations still live in n8n (feedback agent, Slack publishing); Inngest carries the durable scheduled and webhook-driven work on tur-automations. See Toolstack for how the pieces connect.

How it was adopted

Automations run in production for this portfolio and The Unnamed Roads properties. Translation and publishing paths include explicit validation; customer-facing copy still passes human review. Telegram approval is the default for actions that change live systems.

Effect

• Translation: operational run compressed from a multi-week manual pass to an automated first pass (specialist review still required for sensitive copy).
• Publishing: sub-minute path from Slack to live URL for structured posts.
• Inngest workflows: recurring digests and monitors run without manual cron babysitting; site events get acknowledged and retried reliably.

Where AI fits now

Deterministic steps stay in code; models draft and classify; agents (with MCP and approvals) handle multi-step research and coding. This case is the bridge between one-off scripts and agent systems with guardrails — see The Unnamed Roads.