One of the hardest unsolved problems in
enterprise agentic AI is that all the data
these agents need to read was written for
humans not machines.
"john's service is down again, same as
last tuesday" makes perfect sense to anyone
on that team. An AI agent has to guess at
who john is, which service he owns, what
last tuesday refers to, and whether this
is a recurring pattern.
I've been working on this problem — built
an open source semantic normalization layer
that resolves entities, maps relationships,
and normalizes temporal references before
data reaches AI agents.
Tested on real kubernetes/kubernetes GitHub
issues: 99% entity resolution, 15/15 root
causes extracted, 64% token reduction vs
passing raw text directly.
Happy to share more if anyone is working
on similar problems.
1 comment
[ 3.7 ms ] story [ 17.8 ms ] thread"john's service is down again, same as last tuesday" makes perfect sense to anyone on that team. An AI agent has to guess at who john is, which service he owns, what last tuesday refers to, and whether this is a recurring pattern.
I've been working on this problem — built an open source semantic normalization layer that resolves entities, maps relationships, and normalizes temporal references before data reaches AI agents.
Tested on real kubernetes/kubernetes GitHub issues: 99% entity resolution, 15/15 root causes extracted, 64% token reduction vs passing raw text directly.
Happy to share more if anyone is working on similar problems.