I build AI systems that reach production.

The problem

An AI creative tool was shipping, but nobody could answer the two questions that decide whether an AI product survives: what does each generation actually cost, and what happens when a provider fails at 2am?

What I built

A single AI gateway. Every outbound call to every provider goes through one place — per-provider timeouts, structured logging, anomaly detection at 40% error rate or 45s latency, and the ability to cut off a provider through a feature flag, without deploying. Twelve egress points migrated, zero direct calls left.

Then per-node telemetry: cost, duration and tokens for each of the 19 node types in the pipeline. Not aggregate — per step. That is what turns “we have an AI feature” into “we know what it costs and what to charge for it.”

Twelve egress points, one path out

Also shipped

Monthly credit windows with a lifetime ledger as the source of truth. A queue that detects an orphaned job in about two minutes and reprocesses it — without charging the student for the failure. Feature flags with a kill switch. Observable deploys: one endpoint answers which commit is live.

The problem

A post-operative physiotherapy practice running sales, student management and support across WhatsApp, spreadsheets, and an unofficial API that collapsed under load.

What I built

A full CRM with four AI agents that have explicit data boundaries between them: one sees internal operations, one handles inbound leads around the clock, one is isolated and never receives operational data, one writes content.

The sales agent runs on non-negotiable rules encoded in the system, not in a prompt: never offers or negotiates a discount, never closes a sale, never sends a checkout link, never discards a lead. It qualifies and hands off to a human.

Then I migrated the operation to Meta’s official WhatsApp Cloud API — including the diagnosis the documentation doesn’t cover: verifying the webhook and subscribing to events isn’t enough. The app has to be subscribed on the WABA, and the number has to be registered on the Cloud. Before that, Meta simply delivers nothing.

Four agents, four data boundaries

Also shipped

A financial module — amounts in cents, installments, per-seller commission, partner profit split, refunds, event-level audit trail. Approved sales post automatically, deduplicated by transaction. Round-robin lead distribution. Automated first contact with a silent fallback when a template isn’t approved yet, so nobody is left without a message.

The problem

A film production company and equipment rental house tracking a R$519,000 inventory in spreadsheets — and needing it to work on location, where there is no reliable network.

What I built

An offline-first PWA. Equipment check-in and check-out by scan works with no connection: each scan enters a local queue and syncs later without duplicating. It runs on a Mac Mini that travels to shoots, powered by a stationary battery and Starlink.

Autonomous agents with autonomy graded per action, not per agent — A0 through A4. Anything that reaches a client, moves money or deletes data is fixed at A2: the agent prepares the whole action, a human approves, code executes. Everything else moves on evidence — ten clean runs promote an action one level, one correction demotes it and the reason is logged. Every agent action is audited.

The autonomy scale, level by level

Found along the way

Cross-referencing the catalogue against real inventory surfaced 14 items being sold that didn’t exist in the fleet.

The problem

A platform where content creators publish, get measured and get ranked. Every submission has to be analysed, scored and folded into a cycle — in the background, never inside the request. The obvious answer was a queue in Redis.

What I built

The state change and the event are written in the same transaction — the outbox pattern. A worker claims rows with for update skip locked, and consumers are idempotent by key, so the same event arriving twice costs nothing.

Webhooks and server actions record the event and return. Heavy work belongs to whoever consumes it, never to whoever receives the request. Events are named in the past tense — content collected, metric updated, cycle closed — because an event is a fact that already happened, not an order to obey.

One transaction, or two systems that can disagree

The decision that mattered

Queueing outside the database splits the content write from the job enqueue: two writes that can diverge. If the second one fails, the work exists and is never processed — the creator loses points and the ranking lies, silently, in one row. The canonical fix is the outbox, and an outbox table is a queue in the database. Redis would add a piece without removing one.

the argument was never
throughput. it was atomicity.

Where Redis would earn its place

As a rate limiter shared across serverless invocations. So the limiter was born behind an interface, with a PostgreSQL implementation. Swapping it later is one file — when measurement asks, not when intuition does.