Every project below is labeled by what it actually is, and every number is tagged by how we know it. That's unusual, and it's deliberate. AI automation case studies with the provenance printed on each one.
01
Most agency portfolios don’t tell you which projects were paid client work, which were internal, and which numbers were measured rather than estimated. That leaves you to discount all of it, which is the rational response.
Measured means observed after launch and client-confirmed. Self-measured means observed on our own systems, before client confirmation. Projected means modeled and not yet observed.
Labelling the weaker assets is what makes the stronger ones believable.
02
Our own CRM pipeline: definitions before sync, dedup before write, reconciliation after every run, and a dead-letter path for anything that fails anyway.
64 nodes
Every parameter mapping and connection across all four workflows inspected line-by-line, and every code node syntax-compiled
A concept build that makes the eval-first method inspectable: six static checks on an n8n export, a quality bar declared before results exist, regression diffs between versions, capped and labelled AI analysis — and a human gate on what enters the graded suite.
6 checks
Deterministic and static — every failure names the exact nodes responsible, before any AI opinion is allowed
Our own lead pipeline, self-hosted on n8n: malformed rows refused with a reason, retries made safe by idempotent capture, failures escalated to a human — and no number published before the measurement log holds it.
48/48
Field-by-field audit of the live instance against version control — run twice, byte-identical both times, including once after a live editing session
Thirty minutes with the engineer who'd own the build. No pitch, no deck.
Book a build review