Already trusted in the toughest rooms.

SIT brought us a gap every medical school knows about. Their students can name the disease, list the symptoms and prescribe the right drug — and then freeze in front of an actual patient. Nobody had taught them how totalk: how to open, how to ask the question that surfaces the detail the patient did not think was important, how to get the history out of someone who is anxious and vague.

So we built the patient. A 3D avatar sits across from the student and behaves like a real one — volunteering little, hedging, drifting off the point, getting uneasy when pushed too fast. The student has to run the consultation themselves. There is no menu of options and no right answer to click.

The part SIT liked most happens when the consultation ends. The avatar changes character on the spot — same session, new person — and becomes the teacher. It walks back through what the student actually asked, what they could have asked instead, and, most usefully, the questions they never asked at all. Feedback lands while the conversation is still fresh, from the “patient” who just lived through it.

MSF had the same problem in a harder room. Their officers meet elderly residents who cannot follow the process, and residents who have simply decided the answer is no — and the officers were being sent in with policy knowledge and nothing else. You cannot rehearse that against a colleague who is being polite to you.

So the avatar is angry. It sits across from the officer refusing to cooperate, repeating itself, getting louder when it is talked over, and it does not calm down until the officer actually does the thing that would calm a person down. Because the AI is driving the face as well as the words, the officer is reading a real expression while they work — which is exactly the signal they need to learn to read.

Yellow Ribbon runs it for the job waiting on the other side of the gate. An inmate preparing for retail work stands in front of a customer who is furious about a refund — and has to bring that person down. The avatar argues, interrupts, escalates when handled badly, and only settles when the right things are said and done in the right order.

The scenario asks for specific actions, not just soothing words: acknowledge the complaint, check the receipt, explain the policy without hiding behind it, offer the remedy. Getting the sequence wrong keeps the customer angry, which is the honest outcome and the reason the practice works.

One level up from every other engagement on this page. AIC does not want us to train their officers — they want us to train the people who train their officers. The trainers are the ones who have to judge whether an officer handled someone well, and that judgement is the thing that is hard to teach and hard to keep consistent across a cohort.

So the avatar becomes the shared reference. A trainer runs the difficult client themselves, sees what the scenario is actually testing, and then takes it to their own officers with the same standard in their head. The same character, held the same way, in every room.

A course we are writing with the college rather than a deployment we are selling to it. Retail and hotel management students are walking into an industry where the first voice a guest hears is increasingly not a person, and the honest thing to teach them is both halves of that: what voice AI is already changing about the front desk and the shop floor, and how the work that stays human changes shape around it.

The second half is the part that matters. They do not just study the technology, they build with it — students design and stand up their own voice agent for a booking desk or a returns counter, which is a far better way to understand what these things can and cannot do than being shown one.

Four products, one conversation engine — and the conversation engine is ours. The hiring journey is one long conversation seen from four sides of the table, so the same engine faces the recruiter, the candidate, the student and the institution.

On the phone, for the recruiter. Hiring OS screens at a volume no team could staff, and that is our calling stack doing it: already talking as the candidate says hello, stopping when they cut in, finishing the lines that have to be said in full, hanging up on voicemail instead of interviewing a recording, and writing the verdict back as fields rather than a transcript. A hundred first-round conversations instead of the twenty someone had time for.

In the interview. Training OS for the student and Career OS for the candidate both put a live interview in front of a person, and it is the same practice engine as the medical and frontline training above, pointed at a job interview instead of a consultation — questions that adapt, follow-ups that push when an answer sounds thin, under 500ms before it replies, and feedback naming the exact moment it started going wrong with the recording to prove it. Candidates stop performing for it after about a minute, which is the whole point.

See how the four fit together

SilverWings XR is both a customer and the partner we delivered the public-sector work alongside — the training deployments above were built with them. They also run the Voxio Agents voicebot and 3D avatar in front of their own users.

A Singapore client running Voxio Agents as the conversational layer inside their own product — the same engine as the training deployments, without the Voxio Agents name on it.

Every website help system does the same unhelpful thing: it tells you which button to press. If you could find the button you would not have asked.

Ours takes the wheel. You say what you came to do and the agent goes and does it on the page in front of you — scrolling to the right place, opening the right thing, filling what needs filling — narrating as it goes, so you learn the route rather than just arriving. And if showing the problem is easier than describing it, it can look through your camera while you talk.

It is built, it works, and it is running on this site right now.Try it