Arcanum

Precise systems,
quietly built.

Ultra-low-latency infrastructure and applied ML. Two products, measured honestly: what is built, what is measured, and what is still a target.

Rune

A replicated, ordered log with majority commits, transactions, consumer groups, typed topics and materialized views.

Raft-style consensus, acks=all majority commit, partition failover
Thread-per-core broker over Aeron UDP with GSO/GRO; client↔broker TLS 1.3
3 nodes, acks=all, 1 KiB: ≈160–200k rec/sshared 4-vCPU dev box, all nodes co-located — not a tuned host

Surfaces repeated evidence of online-grooming behaviour in chat, with the messages behind it, for a human moderator to review.

Two-tier cascade on commodity CPU: µs-scale n-gram screen, fine-tuned SmolLM2 judge
Flags a user only on repeated evidence; verdicts are signed and carry their evidence
PAN12 test set: precision 1.000, recall 0.622adult-decoy corpus — real-world accuracy will be lower

Every number names its conditions. Targets are labelled as targets until they are validated.

Shared-nothing threads, ring buffers between them, and no allocation on the hot path.

Where software touches people's safety, it supports human judgement. It does not replace it.