Capabilities · Case studies
Case studies
Tests we ran on real builds, with the numbers as measured. Each one is dated the day it ran and links to what it found.
Asking questions across meetings, without a model writing the answer
“What do I still owe?” answered from three meetings in milliseconds, each line cited to its meeting and second. And how the free and paid builds stay apart.
Is this the right conversation? Teaching a small model to say no
Asked whether a lawyer's call was a shop-counter enquiry, a 3B model said yes and filed the case name as the product. Four ways of asking, 22 pairs each, on a laptop CPU.
We installed our own early-access build and tried to make it leak
26 of 26 automated checks passed, and 509 samples of the app's network connections all stayed on the machine. Then we found three things to fix.
Our notes invented decisions. Here is how we found it and stopped it
A phone-store memo came back with “Decided: then let's ship on Friday”. Nobody said it. The words came from our own prompt.
A 20-second voice memo from a phone counter, turned into a cited stock record
A store's demand sheet, filled from a counter memo with no typing: eight fields, each pointing at the second it was said.
A lawyer's client call, filed into the firm's matter log
The filing deadline, the task, who owns it and when it is due: pulled from a 46-second call and cited. And one name the speech model got wrong.
Six Indian-language meetings: Sarvam's cloud models against our on-device pipeline
Saaras v4 averaged 5.2% word error across six meetings. The Whisper models we ship today averaged around 80%. Here is the full run, its cost, and its limits.