We build in the open. The fastest way to judge an AI shop is to look at something it actually shipped and check the work yourself. So here's ours: public, reproducible, and yours to poke holes in. Our edge is empirical model selection paired with auditable, source-traceable output, applied to systems we actually ship and operate. You can verify all three below.
A reproducible pipeline that ingests a document corpus, preserves every source byte with hash-pinned provenance, OCRs every page through a benchmarked vision pipeline (a 9-model bake-off, published error rates), and makes it all semantically searchable with a citing assistant. Open-source, honest about its own limitations.
Read the methodology →
When a new AI model ships, most takes are one run and a vibe. We run a real benchmark panel against our own production fleet: same fixtures, repeated runs, confidence intervals. We publish the numbers with their uncertainty attached. A BPS team publication.
Read Error Bars →
Aurora is our meeting-intelligence product. Instead of piping every meeting through one model and hoping, it benchmarks 14 models across summary, action items, key points, sentiment, and items of interest, then routes each task to the one that measurably wins it, citing every claim back to the exact moment in the transcript it came from. Aurora is in beta access now.
See the methodology →