How we think about meeting forensics, what we ship, and what we will not.
We publish when we have something concrete to say — a system we shipped, an architectural decision we defended, or a piece of methodology we want clients (and skeptics) to understand. No hot takes, no engagement bait, no AI-generated filler. If a post does not survive a re-read a week later, it does not ship.
Transcripts are a commodity. The hard part is what comes after: attributing a sentence to a person, scoring contradictions against earlier claims, and producing something a judge or a board can verify. Here is why we drew that line, and why we will not cross it.
Eight nodes, one Ed25519 signature, and a public key anyone can fetch. A walkthrough of how a Felarity session report is bound to the audio, transcript, and analysis that produced it — and what that buys you in a deposition.
Whisper hears the words. pyannote separates the voices. DeBERTa-v3 re-scores the contradictions on CPU. None of these models is interesting alone. The interesting part is the wiring between them — and the failure modes we had to design around.
No training on customer audio. No selling of voice samples. No third-party analytics on session data. The list of "nevers" is short, specific, and contractually binding. Here is the full list, and the architectural choices that enforce it.
Daubert asks whether a methodology is testable, peer-reviewed, has known error rates, and is generally accepted. We do not pretend Felarity meets that bar autonomously. We do think the attestation chain and per-component error rates make it a useful exhibit when prepared by counsel. Here is how we think about that.
/blog/feed.xml and an email list you can join by writing to [email protected] with the subject "subscribe". We will add you, send you posts as plain text when they go up, and remove you the moment you ask. That is the entire system.