Production AI Engineering

AI systems

production.

Every output traces to the evidence that produced it. A human approves anything irreversible. The pipeline heals itself when it breaks. Because “the model said so” is not an answer your customers will accept.

01

Four commitments, each with evidence

Not values.

These four things show up in every system we have shipped, across two clients and three products. They were never planned as a brand. They are just how the work got done — which is what makes them worth reading.

02

Shipped work

OURS

ColdDeck

cold email that reads like you actually looked.

Finds businesses in a niche, crawls each site for concrete defects, and scores them on a deterministic rule pack — explicitly not a model. Every point traces to the page that produced it. A human approves every email before it sends.

Rule-pack scoringPolite crawlerLocal DBGmail send

Measured

0
MODEL CALLS IN SCORING
100%
HUMAN APPROVED
$0
SAAS SPEND
OURS

DripDeck

two clocks, and neither one needs you awake.

Uploads early and private with a publishAt timestamp, so the platform itself flips the video public at the right moment. Days the machine was off self-heal — the next run notices the missed slot and reassigns it.

SchedulerSelf-healing queueYouTube APITikTok API

Measured

2
INDEPENDENT CLOCKS
0
SLOTS SILENTLY DROPPED
UNATTENDED DAYS

Two further systems — multimodal extraction with end-to-end provenance, and document extraction with auditable confidence scoring — are client work and are not published.

Five days from now

what breaks, why, and what fixing it costs.