Jesse Anderson

Pure-Rust path from cargo build to a working XLL add-in. Handles the Excel C API marshalling, function registration, and the calling-convention quirks that make XLL development painful in the first place. Companion to xllgen (which scaffolds projects) and xll-utils (which validates exports).

Sits between cargo build and Excel. Reads PE export tables out of the produced DLL and asserts the symbols match the registration block, so ABI mismatches surface as build errors instead of opaque “function not available” dialogs in Excel.

A classic categories, threads, and replies forum built as a monorepo with a Worker API, a React SPA, and shared types. Cloudflare Access terminates identity at the edge and hands the Worker a signed JWT, so the app never handles passwords. Features include Markdown authoring, thread polls, full-text search, reactions, mentions, subscriptions, unread tracking, drafts, bookmarks, edit history, a report queue with role-based moderation, per-user write quotas, and encrypted backups with tested restore tooling.

Most of the work after launch went into keeping a free-tier edge app fast and correct. D1 bills by rows read rather than rows returned, and Workers cap subrequests per invocation, so query shape, caching, and what runs in parallel are all designed around those limits.

A private data product for managing a job search without handing personal workflow state to a third-party app. The pipeline runs locally on a schedule, keeps a durable SQLite history, exports static dashboard data, ranks roles against a profile, and deploys the built site to Cloudflare Pages behind Cloudflare Access.

The public description is intentionally architecture-level: it covers the ETL, persistence, ranking, deployment, and browser-local saved/applied/notes workflow without naming target companies or exposing job-board implementation details.

The browser version of a problem I automated at UL Solutions, where detecting changes between CAD revisions ran on Python, OpenCV, and PyMuPDF. Here pdf.js extracts the text layer, diff-match-patch produces the edit script, and the result is painted back onto the rendered page so a reviewer sees each change in position on the page.

A pixel diff runs alongside the text diff to catch what the text layer misses, such as moved graphics or redrawn tables. The pdf.js build is pinned and loaded from its own file, which lets the page serve a content security policy without unsafe-inline.

Auditing an inherited workbook usually means clicking through sheets hunting for whatever breaks. This extracts the whole formula surface at once: formula cells per sheet, defined names, hidden sheets, and links out to other workbooks, which is the set that tends to matter before a migration or a rewrite.

The external-references view exists because cross-workbook links are the ones that silently break when files move. The output states its own boundaries for workbook features the current version does not cover.

A fully client-side SQL environment. Users can write queries against embedded sample datasets with all computation happening in the browser via DuckDB-WASM, so there is no backend to run and nothing to pay for in hosting.

Parquet is the artifact format most of the data work here produces, and checking one normally means opening Python. This reads the footer directly through hyparquet and reports what matters when a file looks wrong: how the row groups are laid out, which codec each column uses, and what the column statistics claim about ranges and null counts.

Compression codecs are fetched lazily and cached, so a file that only needs Snappy never pays the download cost for zstd, gzip, and brotli. The file never leaves the browser.

Parses the entered formula into an AST, differentiates it symbolically with respect to each variable, and evaluates the standard first-order propagation. The contribution breakdown reports which input dominates the combined uncertainty, which is usually what someone actually needs before deciding where to spend effort on better measurements.

A first-order expansion is an approximation, so the tool will run the same formula through a Monte Carlo simulation (normal or uniform, up to 500,000 samples) as an independent check on the analytical result. Coverage factors from k = 1 to k = 3 are selectable, and the page documents its own scope and limits.

A consolidated surface for utilities that benefit from being a URL away. Each tool runs entirely in the browser; the heavier ones (DBSCAN visualizer, graph digitizer, AI image detector) compile through wasm-bindgen for performance.

Real-time face detection running entirely in the browser. Upload an image and faces are automatically detected and blurred using a Gaussian filter, useful for privacy protection when sharing photos publicly.

Visualizes the streaming token output from LLM inference in real-time. Shows throughput patterns, latency spikes, and generation speed variations (useful for understanding model behavior and optimizing prompt design).

The kind of calculation that gets done on the back of an envelope before an animal study and then gets re-derived again every time someone needs to spot-check a poster. Front-loading it as a one-form tool removes the unit-and-formula busywork.

Forward calculation: plasma concentration × Vd → total body load. Back-calculation: starting from a current measurement and a known elimination rate, work backwards to estimate the exposure that produced it. Useful for occupational and environmental exposure modeling where the dose is unknown but the biomarker is measured.

Built while reviewing control theory, on the principle that implementing something is a better way to learn it than reading about it. Running four tuning rules against one process makes their trade-offs visible in a way the individual formulas do not.

The simulation keeps the controller sample time separate from the integration step, so sampling effects show up in the response. Anti-windup is switchable between conditional clamping and back-calculation, and the derivative term can act on measurement or on error. Five preset processes cover normalized FOPDT, fast flow, a jacketed tank, a heat exchanger, and composition control, each with load disturbances and measurement noise. Plots cover the process variable, controller output, and open-loop frequency response.

The guided exercises and the shift-click glossary exist because the tool was built to be studied from.

Built on PsychroLib for the ASHRAE property equations, with charting, process modeling, validation, and export split into separate modules. State is entered as dry-bulb plus relative humidity, with either atmospheric pressure or altitude closing the system.

Processes are drawn on the chart: sensible heating and cooling, cooling with dehumidification, humidification, mixing two air streams, and a coil-leaving estimate. Saturation, RH lines, wet-bulb, enthalpy and specific-volume overlays, and a comfort zone each toggle independently.

Property data comes from vendored tables, with the source and its license named on the page. Pressure can be entered in kPa, MPa, bar, or atm, and the whole tool switches between SI and US customary units.

The page runs internal consistency tests across the vendored data and shows the results, so a user can check whether the tables agree with themselves. Two states can be compared side by side, with phase context and charts around the lookup.