WEBVTT

1
00:00:00.385 --> 00:00:04.878 line:33% position:50% size:85% align:center
Your agents already work through Metrum
AI Router. Turn selected traffic

2
00:00:04.936 --> 00:00:08.698 line:33% position:50% size:85% align:center
into training data, bring a fine-tuned
model back to the same model

3
00:00:08.744 --> 00:00:11.008 line:33% position:50% size:85% align:center
group, and measure whether it earns its
place.

4
00:00:14.373 --> 00:00:18.297 line:33% position:50% size:85% align:center
A support agent looks up a ticket,
checks an order, and finishes with

5
00:00:18.355 --> 00:00:22.965 line:33% position:50% size:85% align:center
tool results. Every model turn goes
through the router, under its caller

6
00:00:23.034 --> 00:00:24.033 line:33% position:50% size:85% align:center
key and project.

7
00:00:27.385 --> 00:00:31.681 line:33% position:50% size:85% align:center
Choose which traffic becomes training
material. Enable governed capture

8
00:00:31.715 --> 00:00:36.568 line:33% position:50% size:85% align:center
for the support project, keep the IT
project uncaptured, and attach task

9
00:00:36.626 --> 00:00:40.539 line:33% position:50% size:85% align:center
outcomes. The router keeps those traces
connected to their caller,

10
00:00:40.585 --> 00:00:42.594 line:33% position:50% size:85% align:center
client, model group, and purpose.

11
00:00:46.507 --> 00:00:51.242 line:33% position:50% size:85% align:center
Export from the router for a specific
purpose, client, project, or key.

12
00:00:51.521 --> 00:00:56.000 line:33% position:50% size:85% align:center
Select successful tasks and a quality
threshold, preserve redaction, and

13
00:00:56.068 --> 00:01:00.345 line:33% position:50% size:85% align:center
carry provenance into the dataset.
Freeze a time-based held-out split

14
00:01:00.401 --> 00:01:04.456 line:33% position:50% size:85% align:center
before training, so evaluation comes
from captured work the candidate

15
00:01:04.478 --> 00:01:05.316 line:33% position:50% size:85% align:center
has not learned.

16
00:01:06.383 --> 00:01:10.673 line:33% position:50% size:85% align:center
Fine-tune on the router-exported
training split. Follow the NVIDIA and

17
00:01:10.760 --> 00:01:14.973 line:33% position:50% size:85% align:center
AMD paths from the same governed
dataset, then return each served

18
00:01:15.017 --> 00:01:18.938 line:33% position:50% size:85% align:center
candidate to the router. Hardware can
change; the application's

19
00:01:19.003 --> 00:01:21.306 line:33% position:50% size:85% align:center
model-group contract stays in one place.

20
00:01:22.571 --> 00:01:26.541 line:33% position:50% size:85% align:center
Register the candidate in the same
support-agent group. Canary it only

21
00:01:26.599 --> 00:01:30.651 line:33% position:50% size:85% align:center
for the support key, with budgets and
quotas. Other callers keep the

22
00:01:30.686 --> 00:01:34.784 line:33% position:50% size:85% align:center
teacher. The router controls exposure
without changing the agent's model

23
00:01:34.842 --> 00:01:38.116 line:33% position:50% size:85% align:center
name, and keeps rollback available at
that same boundary.

24
00:01:42.560 --> 00:01:46.552 line:33% position:50% size:85% align:center
Replay the frozen held-out router traces
through the router. Compare

25
00:01:46.608 --> 00:01:50.457 line:33% position:50% size:85% align:center
teacher, base, and candidate on the same
tasks, and check general

26
00:01:50.513 --> 00:01:55.532 line:33% position:50% size:85% align:center
regressions. Measure successful
completion, latency, and cost, then let

27
00:01:55.576 --> 00:01:58.665 line:33% position:50% size:85% align:center
the preregistered gate decide whether
this caller should receive the

28
00:01:58.698 --> 00:01:59.327 line:33% position:50% size:85% align:center
candidate.

29
00:02:00.385 --> 00:02:05.017 line:33% position:50% size:85% align:center
Both candidates fail the promotion gate.
Neither earned promotion. Keep

30
00:02:05.064 --> 00:02:09.313 line:33% position:50% size:85% align:center
the teacher in place. The router governs
the loop, and stops models that

31
00:02:09.348 --> 00:02:10.869 line:33% position:50% size:85% align:center
do not earn the right to ship.

32
00:02:15.385 --> 00:02:19.030 line:33% position:50% size:85% align:center
Start with your agents' work in the
router. Build a governed learning

33
00:02:19.088 --> 00:02:21.468 line:33% position:50% size:85% align:center
loop, and keep control of what ships.
