Case studies

From a simple workflow to a full AI agent — the range of what we build.

Two engagements, chosen to show the range: a workflow automation for a small business, and an AI agent for a mid-size company. We don't publish client names without permission, so engagements are described by their shape rather than their client.

Small business · Workflow automation · Delivered on self-hosted n8n

Consolidating revenue across five ad networks

The problem

Revenue from five ad networks (AdMob, AdSense, Unity Ads, AdColony, AppLovin) had to be consolidated by hand every week — different formats, currencies, and identifiers per network. One cycle took half a day.

What we built

A scheduled workflow that normalizes all five exports, converts currency using each row's own date, overwrites restated figures instead of duplicating them, and flags any network that didn't report.

The result

The weekly reporting cycle went from about 4 hours of manual work to roughly 30 minutes of reviewing flagged items — around 3.5 hours recovered every week. Restatements are absorbed automatically instead of being corrected by hand a week later, and a network that fails to report is caught before the numbers are published rather than showing up as a bad week.

Mid-size & enterprise · AI agent · Custom Engineering

An AI agent that triages every support ticket before a human sees it

The problem

A 50-person software company fielded around 200 support tickets a week, all triaged by hand — read the ticket, look up the account, check the knowledge base, then route or escalate. First response averaged four hours.

What we built

An AI agent that reads every ticket on arrival, pulls account and usage data, and classifies intent and urgency. High-confidence questions get an instant, cited answer; everything else routes to the right queue with context attached, and churn-risk signals go straight to the account manager.

The result

38% of tickets are now resolved with no human touch, and first response on high-confidence tickets dropped from about four hours to under five minutes. Everything the agent can't close arrives at the right queue already enriched, so the humans start from context instead of from scratch.

Tell us what's still manual.

Start a project →