AI knows our words. Let’s teach it to read the room.
We bring the signal,
not the verdict.
Borry measures voice, face and body with about 200 open models and the formulas of behavioural science. It gives those signals to an LLM alongside the transcript. The LLM does the interpreting.
Watch the demoMario talks to Claude. Words alone, then words + Borry.
The responses from Claude are real.
93-second demo · Sound on
There’s more to a conversation than its transcript.
A mother sees a difficult day at school in a glance. An experienced doctor notices what a patient leaves unsaid. People read posture, expression and voice alongside the words.
These are patterns described by behavioural science. An LLM given only a transcript misses those signals. Borry makes them available as measured data.
From a recording to evidence.
-
Measure what happened.
A meeting, a call or a video. About 200 open audio and video models extract the raw signals.
Voice and prosody, volume, laughter and hesitations, speaking turns, facial action units from multiple FACS detectors, gaze, and head and body movements.
-
Put the science to work.
Our research engine has read 3,252 papers in behavioural science and extracted 2,331 formulas.
Each formula is stored as data, with its source and page. It is calculated on the recording and labelled direct, approximate or hint.
-
Let the LLM interpret.
The LLM reads the measurements and formula results. It describes what happened and when, citing the exact data row behind each claim.
We bring the signal, not the verdict.
What changed
in the first test.
The same 50-second video. The same model. Three analyses per condition, compared by a blind judge: transcript only versus the full set of Borry data.
| Measure | Transcript only | With Borry |
|---|---|---|
| Claims per analysis | 33 | 72 |
| Claims with a measurement | 8% | 59% |
| Claims with a timestamp | 37% | 43% |
| Claims supported by the data | 96% | 96% |
| Preferred by the blind judge | 0/6 | 6/6 |
One video, one person; tests on many public videos and on labelled datasets are next.
For AI coaches, therapy support, robotics, companions and research. And for the data models learn from: multimodal, measured, structured.
We’ve been
building together.
Calogero MandracchiaRaffaele NicodemoCo-founders, Borry
We met in 2016 at Deloitte Italy’s Blockchain Lab. Since then, we’ve founded:
- Efebia, 2018. Our first company, from blockchain startup to software house and startup incubator.
- Zipaki. The “Spotify of Italian comics”.
- HibeTV. A “Netflix” where celebrities and their fans are the stars.
- Deepen. Long-form articles read by professional voice actors.
- Terminus. A US foundation for an anonymous voting system built on a fork of Monero.