Building the symbolic representation layer for AI.
San Francisco, CA
Joined February 2026
- Tweets13
- Following3
- Followers13
- Likes16
Agentic harnesses are making AI systems more capable. They're also making reproducibility a lot harder. #AgenticAI #Reproducibility
Can't reproduce a decision? Then you can't tell if the next version improved, regressed, or just changed.
Everyone's talking about slowing AI down.
But how do you control a system that can't even repeat its own behavior twice?
Once AI is making decisions inside real products, that stops being academic. It's a debugging problem, a regression problem, a compliance problem, eventually a trust problem.
The SEMQ Group retweeted
A few months ago, mostly out of curiosity, I started exploring whether embeddings could be reduced purely to lower storage costs.
The initial idea was simple: what survives if, instead of representing embeddings by magnitude, you represent them through a directional pattern?
After many experiments, something even deeper showed up.
Semantics wasn’t in the magnitude at all but it lived in the directional structure.
I verified this wasn’t classical quantization. What emerged was fundamentally different.
On January 3rd, we presented what we now believe points to a new representation layer for embeddings:
it compresses, eliminates drift, and enables stable semantic identity.
Whitepaper/IR: zenodo.org/records/18141590
Landing: semq.mindustries.tech