Kyle Weinstein

An equity research pipeline where every factual claim traces to a specific, dated, tier-labeled primary source. Information sourcing and citation verification are enforced deterministically by code wherever possible to minimize reliance on LLMs, reducing both cost and hallucination risk.

• Seven-tier source hierarchy from SEC filings down to unaudited public posts, with market data marked as a point-in-time price fetched at the time of the research.
• Claims are extracted by regex, not by a model. The citation grammar is generated from one module the pipeline and the prompt both import, so they cannot drift apart — the code raises at import if they do.
• Mechanical tie-out with a digit guard. A quoted figure is located in the source by exact search; a one-digit change leaves a ~97%-similar string, so similarity alone would confirm a fabrication. Comparing digits does not.
• The audit separates proven from judged instead of averaging them into one reassuring number, and deep-links each proven claim to the exact sentence in the filing.

The same guarantee, built as a human-first workflow for tracking investment theses, receiving a feed of thesis-pertinent snippets directly from SEC filings, and recording how each snippet supports or contradicts the thesis. It is designed for source-adaptability, cost-efficiency, and auditability, so that every investment decision can be defended by a verifiable chain of evidence.

• Built against customer research. A family-office research lead's line — "tie it out or it doesn't go in" — is the specification the code implements, not a slogan added afterward.
• Uncited claims are unrepresentable. Structured output forces every candidate claim to carry a verbatim excerpt and a source pointer at the moment it is drafted, when the writer still knows what it was looking at.
• Two independent checks per claim — mechanical substring tie-out, then a separate model pass judging whether the located excerpt actually supports the claim as written. It catches the real-excerpt, overreaching-claim case.
• Detect and quarantine, never silently fix. Failed claims are excluded from the report body and surfaced with their reason; time-to-draft and cost are stamped into every report.

An interactive laboratory for designing, stress-checking, pairing, and rebalancing long-only portfolios from adjusted daily-close history. It keeps the analytical model transparent while making dense portfolios fast enough to explore directly in the browser.

• One diversification map scales continuously. A clustered Canvas explorer moves from portfolio overview to exact pair values through semantic zoom, search, pan, and keyboard navigation instead of switching between portfolio-size-specific designs.
• Large analyses stay off the interface thread. Correlation, classification, clustering, and constrained optimization run in a Web Worker over packed typed arrays, keeping draft editing responsive as holdings grow.
• Pairwise work is reduced to O(N²T). Return series are aligned and standardized once, the upper correlation triangle is stored once, and dependent tools reuse the same matrix rather than rebuilding statistics for every cell.
• Drafts and analyzed results are deliberately separate. Expensive work runs only when requested, prior results remain visible while inputs change, and stale requests cannot overwrite a newer portfolio snapshot.