Services
The whole arc, from the first conversation to a working build, done by the same pair of hands so nothing is lost in a handoff. Hire for one part or the whole span. Every engagement starts with the real problem, not a scope document.
Jump to: Strategic UX research · Service and interaction design · Facilitation and design leadership · Agentic platform engineering · Label Test
Strategic UX research
Talk to the people who actually use the thing. Turn what they say into decisions a team can act on. Interviews, field studies, journey research, and the synthesis that makes it usable by the people who have to build it.
Service and interaction design
Map the whole journey, not just the screen. Design the system underneath it: the data, the hand-offs, the business rules, and the interface that makes all of it feel obvious. See how the work gets made in the design examples.
Facilitation and design leadership
Run the sprint, align the stakeholders, get a room full of opinions to a decision. Workshops, design sprints, and the steady leadership that keeps a product team pointed at the same outcome.
Agentic platform engineering
The new part. We no longer just hand over research and wireframes. We ship the production software too, built one person deep with agentic AI and a human deciding what is true, what is safe, and what is worth shipping. That now includes small custom models: tuned on your domain, checked before they ship, and run on hardware you control.
Label Test
A private service for teams with LoRA or LLM data who want people, not only models, to check it. Training data, model answers, chat transcripts, generated video: if you need to know which ones are right, which ones make things up, and which ones just look wrong, Label Test puts them in front of people through a shared link.
How it works:
- The judges go first. A panel of open-weights judges from different AI families reads every record and estimates how often each judge gets it wrong. Most records are settled there.
- People settle the rest. What the judges cannot agree on, plus random spot checks of the judges themselves, goes to a private test at label-test.humanjava.com. You decide who gets the invite link: your team, your customers, or reviewers we recruit.
- Five at a time, blind. Reviewers see a few items, never which model produced them or what the judges said, and need no account. They leave a name or initials and their labels, never an email or IP address (privacy).
- You get it all back. Every record with its verdict, the judges' error rates, and what your people said, ready to keep, fix, or drop before you train.
We open a test when you need one and close it when the answers are in. No closed or API model touches your data; the judging runs on our own hardware in British Columbia.
We built it to check our own cooking model: four judges agreed it went from inventing recipe details on 25.8% of questions it could not answer to 2.7%. If you are training your own model and want that kind of check, let's talk.