Selected Work — Rise Solutions LLC

Case studies

A brand voice we were licensed to record but not to improvise

Problem: A product feature needed a live, conversational voice assistant. Our established brand voice was the obvious choice and the one users already knew.
Constraint: Our vendor's terms permit producing fixed audio assets but prohibit live synthesis inside a hosted product. Ignoring that would have put a commercial agreement at risk to save a design decision.
Decision: We split the role in two rather than work around the licence. The brand voice became a one-way broadcast character — scripted, produced in advance, and narratively incapable of conversation. A second voice, on a vendor licensed for live use, handles the interactive half.
Outcome: The limitation became the characterisation. Pre-produced audio also costs nothing per play, so the atmospheric layer stays free for every user while only the interactive layer draws metered credits.

A product that was structurally English-only, and nobody had noticed

Problem: We planned to localise a text-input product into many languages.
Constraint: Testing revealed that input normalisation discarded every character outside the Latin alphabet. Correct answers in Hindi, Russian, Chinese, Japanese and Greek all reduced to an empty string — indistinguishable from each other and from wrong answers. Accented Spanish degraded silently.
Decision: We treated it as an architecture problem rather than a text bug. The matching logic moves server-side, where it can be Unicode-aware, hold per-language alternatives, and tolerate the variation speech recognition introduces.
Outcome: One change unlocked four things we had been scheduling separately: multilingual input, voice input, verified ownership of user-earned items, and real anti-cheat. Found before launch rather than after.

Selling small digital goods across tax jurisdictions

Problem: Low-value digital purchases sold internationally. Registering for and filing consumption tax in every jurisdiction is expensive, ongoing, and in some countries required from the very first sale.
Constraint: Selling through a platform that acts as reseller shifts that burden — but costs roughly a fifth of revenue, and restricts what may be sold and where.
Decision: We priced both routes against each other instead of picking by instinct, and derived a break-even in annual revenue per market. Below it, the platform is cheaper and carries no compliance burden. Above it, registering directly wins. We also separated the two rails by a clear test: does this purchase consume a real service, or is it purely an entitlement?
Outcome: A written rule for which rail a given product uses, rather than a decision re-argued every time — and no perpetual filing obligations taken on before the revenue justified them.

Verifying 100 locations against authoritative sources

Problem: A product needed accurate coordinates for 100 well-known places.
Constraint: The obvious authority — UNESCO's official inscribed coordinates — publishes property centroids. For large sites those miss the actual monument badly: by 10.6 km at Mesa Verde, 10.5 km at Rapa Nui, 2.9 km at Angkor. One site's own official tourism page published a figure roughly 48 km from the wall it described.
Decision: Primary sources first, but verified against the thing itself rather than trusted by reputation. Where no authority publishes a structure-level figure — five of the hundred — we used surveyed map geometry and recorded the provenance of every value, including which were authoritative and which were not.
Outcome: A dataset whose weakest entries are labelled as weak. An error found later is traceable to its source in seconds.

How we work

Verify before asserting. "Done", "deployed", "fixed" and "verified" require evidence from this session, not memory. Most of the defects above were found by checking a claim that was already believed.
Primary sources first. Vendor terms, official documentation, the API itself — before any summary of them. A model's capability is not the same as a specific configuration's capability, and only one of those is worth building on.
Constraints are design input. The strongest feature in the list above exists because a licence forbade the obvious approach.
Ship reversibly. Work is committed in small steps with a named recovery point. Production, money and anything outward-facing waits for a deliberate decision.

Not shown here

Living Languages

An initiative through Quavril opening our conversational tutors to communities working to keep endangered languages in daily use — community-led, speakers credited and compensated, aligned with the UNESCO Decade of Indigenous Languages. The program is beginning; we are seeking our first partnerships. Read the commitment →

Historical cryptanalysis

Independent research: statistical attacks on undeciphered manuscripts. Calibrated detectors, in-domain controls, six corpora tested — and the nulls reported as nulls. The record includes the instruments that failed calibration and the signals we killed, because in this field that is the part that counts. A paper is in preparation.

Confidential development

Part of our work is unreleased and is not described publicly — including research we intend to protect. It is deliberately absent from this page rather than vaguely alluded to. Under an appropriate agreement we can discuss scope and results.

Shipped products, and the decisions behind them.

Live products