18 projects

Engineering focus

Every route into the product - a purchase, an admin grant, a self-serve refund, a chargeback from the Stripe dashboard - has to end at the same answer about who can see what. So access is decided at a single entitlement gate rather than checked ad hoc per endpoint, and certificate issuance is idempotent from both the lesson-completion and quiz-submission paths, so either ordering mints exactly one certificate.

Engineering focus

Vendor risk tools tend to return a single opaque number. The design question here was how to keep risk and confidence as separate dimensions, so that a thin evidence base lowers confidence rather than quietly inflating or deflating the risk score - and how to make every record point back to the document it came from.

Engineering focus

The central constraint: language models are allowed to classify passages against a closed nine-mechanism vocabulary and verify quotes verbatim, but every numerical score comes from deterministic additive rules over curated modifier cells. This keeps the output reproducible and lets each number be traced to its source passage.

Engineering focus

A time-series model will happily report an excellent score while quietly reading its own answer. Most of the work here went into making that impossible: backward-only causal rolling-median imputation, limited forward-fill windows and shift(1) change-rate baselines, so no feature at time t can observe the target it predicts. The training and inference paths were split into separate feature groups for the same reason - an observed-only historical store and a forecast-weather store - so forecast data can never contaminate training.

Engineering focus

Generic task apps assume the user can decompose work and judge what to do next. The premise here is that they often cannot - so ordering is computed rather than left to the user, by a deterministic four-factor engine scoring weighted priority, deadline urgency, mood-to-effort match and mood-to-task-type match. Keeping that scoring deterministic matters: the model assists with decomposition, but it never decides what you work on, and when it is unavailable the app degrades to manual entry instead of failing. Decomposition itself meant fine-tuning candidates small enough to serve cheaply behind an ordinary mobile client.