I've been shipping software with AI since 2018
In 2018 I co-founded Speak Technologies to build a platform that placed, transcribed, and analyzed voice communication at scale. To make it useful we shipped sentence classification and named-entity recognition on real call traffic. We were training, evaluating, and operating models years before “AI” meant a chat window.
Why does the date matter? Because the AI conversation has developed a strange amnesia. To hear it told, applied AI began in November 2022. But the discipline of making machine learning earn its keep in production (the data pipelines, the evaluation, the gap between a good demo and a dependable system) is older than the boom, and those of us who lived that gap read the current moment a little differently.
What pre-LLM AI taught me
Three lessons from the Speak years that never stopped being true.
The model is a component, not a product. Our NER model was useful only because of everything around it: the call recorder feeding it, the pipeline cleaning transcripts, the review surface where its output met a human. Swap “NER model” for “coding agent” and the sentence still holds. Every word of it.
Evaluation beats intuition. A classifier that feels right and a classifier that measures right are different classifiers (we learned that one the expensive way). We learned to distrust the demo and trust the confusion matrix. Today the equivalent is distrusting the impressive transcript and trusting the merge record: of the agent-written changes that reached production, how many survived review unchanged?
Failure is a distribution, not an event. Production ML fails at the margins, quietly and constantly. You engineer for the failure rate rather than pretending a better model will end it. That mindset is exactly what agent operations demand now, and I’m pretty sure it’s the least transferable skill from demo-land.
So what did LLMs actually change?
The cost. When ChatGPT-class models arrived I integrated them into my own platform early, first for the obvious things, then for process automation inside my company’s client work. The change was real, and it wasn’t magic: the per-capability cost collapsed. Tasks that took Speak months of data work (classify this, extract that, summarize this call) became an API call. Months to minutes. That part deserves the hype.
What didn’t change: everything around the model. The pipeline, the evaluation, the human gate, the failure budget. The companies that got burned by AI these last few years are mostly companies that treated the collapsed cost of capability as a collapsed need for engineering. It wasn’t. It never is.
From models to agents
By 2025 the frontier had moved again, to models that act: write code, run tools, push branches. I went in the way I go into everything, by building the operational scaffolding first (old habits from the radar floor die hard). Today I run a fleet of AI coding agents that builds my products daily, with work queues, scoped credentials, and transport-level guardrails. That discipline grew into a product suite of its own.
But the through-line matters more than any single system. From a 2018 NER pipeline to a 2026 agent fleet, the job has been the same: take a powerful, unreliable component and build the system where it’s reliably useful.
The boring superpower
“AI-native” gets thrown around like adopting the newest model is the achievement. The durable advantage is operational: the judgment to know what machines can own, the engineering to contain what they get wrong, and the humility to keep humans on the gates that matter.
That’s not a 2022 skill. I’ve been practicing it since production NLP was deeply unglamorous, and before that in fields where the discipline came first: radar rooms, test cells, factory floors. The tools finally caught up to the method. From where I’ve stood, that’s the whole story of AI so far (the method was always the product).
I'm Joseph Dattilo — engineer-founder in Lansing, Michigan, author of the FleetHarbor suite, and founder of Date Palm Media. More about me · More writing · Get in touch