Callum Doty

Analytical Systems & Decision Science

I build analytical systems at the intersection of software engineering, statistical modeling, and decision science. My work focuses on translating complex, imperfect data into practical tools for forecasting, optimization, and operational decision-making.

My approach is to start with the decision that needs to be made, identify the variables that matter, construct a representation of the underlying problem, and build a system that turns that representation into something actionable.

I am interested in the process of turning an ambiguous real-world problem into a tractable, computable one.

Selected Projects

  • Quantitative Weather Trading System

    A quantitative trading system for Kalshi's daily high-temperature markets: probabilistic forecasts, a market-facing scoring benchmark, and serverless execution live with real capital across 7 U.S. metros.

  • House Campaign Allocation Model

    Where should a party committee spend its money across House races? A calibrated forecast of every race, an explicit assumption about what party money does, and a test of DCCC's real spending against the best use of the same budget.

  • AI Political Mail Catalog

    An AI document intelligence platform for political campaign materials, using Claude, hybrid semantic search, and automated Dropbox ingestion to make 9,800+ past pieces searchable in plain language.

All projects

Recent Articles

  • The Crash Loop

    My document worker restarted every five minutes with no error in the logs. The cause was a 0.8 KB PDF that rendered to 815 MB, and the loop exposed four bugs in the recovery code built to survive exactly this.

  • The Login That Failed Open

    A sign-in error I blamed on my hosting platform was a middleware ordering bug, and the emergency fallback I wrote around it quietly turned authentication off for three months.

  • Reading the Fine Print

    How my political mail catalog reads each piece's client, state, and date with regular expressions, by ignoring the message and reading the two lines every piece carries for someone else: the legal disclaimer and InDesign's export stamp.

  • Asked for 500, Got 40

    A pgvector search that asked for 500 candidates got 40, and its filters ran after that. Tracing the cap to HNSW's ef_search, why raising it cost 12–21 seconds a query, and the scroll-driven candidate window that shipped instead. Part 3 of 3.

  • Vectors in Postgres

    How my catalog stores 1,536-number embeddings in Postgres with pgvector, why an exact nearest-neighbor search reads every vector, how an HNSW index walks a graph instead, and the keyword search that runs beside it. Part 2 of 3.

All articles

Elsewhere