Speak Technologies

The voice platform I co-founded in 2018 — where I learned to ship AI in production before "AI" meant a chat window.

What it was

Speak Technologies built a platform that placed, transcribed, and analyzed voice communication at scale. A nine-person distributed team shipped the whole stack: call placement and recording, transcription pipelines, and the part that made it genuinely useful — production natural-language processing. Sentence classification and named-entity recognition running on real call traffic, trained and evaluated by us, years before large language models made that look easy (it was not easy).

The NLP was the product, not a demo. Classification and entity extraction had to hold up on messy real-world transcripts — accents, crosstalk, half-finished sentences — where a model that benchmarked well but failed quietly cost customers real money. We built the training data, the evaluation harnesses, and the deployment pipeline ourselves, because in 2018 there was nothing to rent. That end-to-end ownership — data to model to production behavior — is the part I carry forward into everything I build now.

What it taught me

Speak is where my AI story stops being a claim and starts being a date: 2018. The lessons from operating models in production — the model is a component, evaluation beats intuition, failure is a distribution — are the exact discipline I run AI agent fleets with today. I wrote the longer version in "I've been shipping software with AI since 2018."

The company wound down in 2024. Six years, a real product, real customers, and an education in applied AI you couldn't buy at any price (believe me, we tried other ways first). When the board voted to dissolve it, I personally bought all the IP back — not because it made business sense, but because I was proud of what we built together and wanted the legacy to live on in some small way.

Full career context on the track record. Old speakmeetings.com links land here.