Most of what we make is designed and directed by our creative director and built with frontier coding agents. What these pieces share is a habit: decide how the work will be checked before building it.
Evaluation datasetTaller B-0
A parity corpus of 1,000 propositions, each written in Spanish, English, Chinese and Mas+, our own symbolic notation, with its interpreter built in. Five strata, each with its own grader: execution, camera, exact-match QA, testimony, and calibration for the unknowns.
Rows can be flagged private and redacted before the freeze, and the export's hash is published before any training run.
Speech · linguisticsCode Switcher
A text-to-speech prototype that infers who is speaking from how a sentence switches between Spanish, English and Chinese, instead of cloning a voice. Seven stages, from language ID and switch typology to eight speaker dimensions.
A box on the page says which stages are real and which are hand-authored, so the inference gets tested before anyone spends a training run on it.
Forecasting · editorialThe forecast desks
NVIDIAddict, AMDdict, AAPLicted and Cyanide: research desks for single-stock holders. One engine with opposing editorial voices, bull and bear, and a satirical TV desk that predicts what each panelist will say.
Every forecast is sealed, then Brier-scored against the tape, and an independent validator clears each question before it seals.
nvidiaddict.com ↗
CalibrationWongstrology, Observatory edition
Astrology and Arabic letter science, audited and offline. Every claim carries an evidence tier, readings face a blind test after Carlson (1985), and bets are hashed when you commit them.
Moon positions are checked against the Swiss Ephemeris (0.25 arcminutes at worst), and a parity test holds the browser engine to its Python reference.
Science writingDivide by Zero Atlas
Fifteen open problems in science on five shelves (undefined, undecidable, unmeasurable, unreconciled, unaccounted), each with a working model and a status as of October 2026.
It treats "never" with the same suspicion as "known", and flags preprints, folklore quotes and single-sourced facts in a cautions note.
Editorial pipelineLLL+ Publishing Portal
A control room for translated, annotated editions of old texts, with a review agent that logs every run, updates each program's stage and writes the brief for the next session. Some calls are reserved for the human, and the agent can only ask for them.
Its first edition, Shams al-Maʿārif, is rendered from the Ḥalabī 1927 print collated against three Paris manuscripts, in Arabic, English and Spanish.
Brand architectureBirthstone Spirits
A spirits house built on the birthstone calendar: twelve months, twelve stones, two rules, one reserved. Liqueurs wear their stone in the bottle; whiskeys wear it on the band.
It ships with a regulatory ledger: formula approval, protected place names, and which famous trademarks to keep unnamed.
Licensing · agentsCue Router
Music sync licensing as explicit routing: set your use, budget, term and rights, and it scores twelve buyer routes and nine seller routes with stated reasons and a "check before you commit" list.
It shows the request and response an agent tool call would consume, because the next buyer may not be a person.
Multimodal toolingPila
Build an ordered stack of images with your notes, then send it to a model in one pass or image by image: Claude on your own account, a local model, any compatible server, or a webhook.
When the model offers to rewrite a note, you accept it or keep your own. Keys stay on the device.
App shelfMoHuOS Open Shelf
Sixteen single-file apps in Spanish, English and Chinese with no accounts: a stroke-based programming language and its IDE, script atlases, a glyph tool, and a word-dating quiz that answers "unknown" on purpose about one time in five.
Every claim names its source; where nothing is known, the page says unknown.
Ranking designA rubric that scores behavior
A re-ranking rubric for a community app that scores the conversation, never the person: disclosure scores the act, never the content; nothing goes below zero; and a 14-day window means nobody carries a permanent mark.
Weights sit on sliders in an admin panel with a live rank check.
Most pieces are private builds; walkthroughs and screenshots on request.