R/01
Agent design and authority
How an AI agent should reason about its own capabilities, when to defer, and when to refuse. Not "alignment" as a slogan — actual decision policies that keep an autonomous system inside the lane its operator drew.
/ RESEARCH 06 DIRECTIONS DIRECTIONAL · NOT PUBLISHED
What we think about
when we're not shipping.
The areas of AI we find most interesting, and where our internal work sits. Not papers — directions. Some of these become products. Some inform how we build for clients. Some stay in the lab.
R/02
On-premise & privacy-first AI
Most useful AI work today routes private data through someone else's GPUs. We think the inverse should be normal: models run where the data lives, audit trails belong to the owner, and "AI" doesn't mean "cloud dependency."
R/03
Document understanding for small businesses
Contracts, receipts, intake forms, recurring paperwork — the boring middle of every operation. Underrated as an AI surface, because it's where the real time gets lost and where small reliability gains compound fast.
R/04
Multilingual systems in production
Real-time translation, mixed-language workflows, bilingual document generation. Hard problems where the failure modes are subtle — false fluency, missing nuance, lost domain terminology — and where shipping requires more than a benchmark score.
R/05
AI characters and persistent personality
How do you get a model to behave like a consistent person across long arcs — a campaign, a season, a product lifetime — rather than a session-shaped chatbot? The platform layer underneath Anima.
R/06
Interactive narrative and game AI
Story systems where AI participates as a designer's tool, not a substitute for one. Engines, not endpoints. (Patent work in progress.)
Working on something in one of these areas and want to compare notes? We're reachable. Especially happy to talk to academic researchers, defense-of-fundamentals types, and operators who actually have to ship.
Reach out →