LLMs, agents & evaluation

Generation system that turns a single learning goal into a complete training world: knowledge graph, scenarios, stages, simulated app, and rubrics.

Conversational assistant that guides users on what to buy. One of Tokopedia's first generative-AI products.

A voice-and-vision AI agent that sits beside a person while they work on a task on their computer: watching their screen, pointing at what to do next, and checking whether the piece of work is coming together correctly.

Simulation runtime that can stand up any software tool and run an agent or a user through it, with specific understanding checkpoints baked into every task.

A chat system that surrounds a person with a full cast of AI coworkers, a manager, peers, and specialists, each with their own personality and memory, so working through a task feels like working inside a real company.

A support flow a person can step into mid-task when they're stuck: the same AI teammate they've been working with pauses the task and walks them through exactly what they're missing before they go back to it.

A voice agent that talks to someone in their first few minutes on a new platform and builds their profile, goals, prior experience, how they think through problems, from that one conversation instead of a signup form.

A real-time voice flow where a person presents a piece of work out loud to an AI audience, then fields live follow-up questions about the specific things they said, ending in a structured verdict on how they did.

A system that continuously reads how well a person is actually performing, independence, quality, communication, speed, and turns that read into an experience-point score that reflects real skill, not just activity.

A classifier that reads whether a person is confused, stressed, in a state of flow, or checked out from their recent activity, and uses that read to decide whether their AI manager should be more supportive or more challenging in the next conversation.

A running record of the working relationship between a person and their AI manager, so a conversation on day ten feels like a continuation of a relationship instead of a conversation with a stranger.

A submission system that reads whatever a person turns in (a document, image, code file, or link), grades it against a rubric, and catches people who resubmit the same broken work hoping it slips through the second time.

A review that looks across everything a person did across an entire multi-part project, not just one submission, and names the pattern: where they got faster, where they kept stumbling, how they grew.

Two short understanding checks that run right before someone starts a task, to catch a misread of what's actually being asked before it turns into days spent building the wrong thing.

A continuously updated story of how one person is actually progressing, built from every task, conversation, and submission they complete, so the system remembers the pattern instead of re-reading them from scratch every time.

A per-call log of every grading decision the system makes, plus a dashboard that reads that log across many people at once to catch a task that's grading too easily, too harshly, or drifting from what a human would say.

A voice walkthrough where a person's AI manager reads their grading verdict out loud, pointing at the exact spot on their submitted file each comment is about, instead of leaving them to parse a written report alone.