Delta
A filing-change intelligence engine: type a ticker, get a five-year report on what actually changed across a company's 10-Ks, every claim traceable to a deterministic diff.
Delta fetches a company's last five 10-Ks, aligns them section-by-section through a stable anchor scheme, and computes a deterministic paragraph-level diff: cosine-similarity buckets classify each change, numeric deltas are pulled from XBRL structured data, and only the changed passages ever reach the LLM. The governing principle is that the model never finds the diff, it only explains it: every interpretation links back to a diff record with verbatim side-by-side quotes, which keeps the system cheap and structurally hard to hallucinate. Materiality is triaged (boilerplate → notable → material) so real changes surface above churn, and churn scores per section headline each report. Thresholds were tuned on a hand-labeled 50-pair sample; it's built on the v1 RAG eval harness, an 8-config chunking × embedding × rerank matrix where section-aware chunking beat fixed-size 28/56 vs 9/56, retained as a regression check. Server-rendered with FastAPI + Jinja2, deployed on Fly.io.
Diff any ticker @ Delta
Visit Relay ↗