Not Sure Where to Start With Musubi? Here's an Honest Guide

The short version of what we do

Musubi builds building blocks for Safety and Fraud teams shipping with AI: custom models, agents, workflows, policies, and observability, knit into one system that learns from your team. It comes down to two things: spotting harm (observability) and then doing something about it (action). Everything is modular and mix-and-match, so you can start with one piece, and it's most powerful when you combine them.

Today that's these building blocks:

• Content Moderation — LLM-powered content moderation that applies your policies consistently across text, images, and other user-generated content. We give you a variety of models to choose from and include eval tools, policy templates, and more.
• Fraud & Spam Enforcement — Custom-trained enforcement models that detect fraud, spam, and other platform abuse that learns from your moderators and adapts daily.
• Real-Time Chat Moderation — label live messages in under 150ms, steerable like an LLM but priced like a classifier.
• Content Insights & Analytics — A reporting and analytics layer that helps teams understand moderation trends, emerging risks, and policy performance over time. Clusters similar content together without needing pre-labeled data.
• Agentic Investigations — hand an AI agent a single signal and get back a full, sourced case for your team to decide.
• Moderation Console — A centralized workspace where moderation teams can review cases, investigate edge cases, and manage enforcement workflows more efficiently. Based on open-source Coop from ROOST.
• Workflows — A flexible workflow engine for building, automating, and scaling trust and safety processes across tools, teams, and use cases.

Most teams start with content moderation or fraud, and grow into the rest. Here's each one in more detail.

Content Moderation: for consistent, explainable enforcement

Great if your main challenge is content (messages, posts, images, listings, usernames, videos) and you need to enforce custom policies at scale.

Instead of training a custom classifier or wrestling with a vendor's pre-built labels, you write your policies the way you already write them. It reads them and gives you what you need to enforce. Update your policy, and enforcement updates immediately. No engineering ticket, no retraining. Your policy or T&S team has full transparency and control.

Every decision comes back with a label, a severity score, a confidence score, and a rationale. That rationale matters more than it sounds, because it's what lets your moderators calibrate the system, and it's what you show a user when they appeal.

A few things worth knowing:

• It handles text, images, audio, video, and multimodal combinations. Evaluating an image and its caption together catches things that evaluating them separately would miss. Text moderation covers 150+ languages (everything frontier LLMs cover).
• You choose your model tier. Faster and cheaper for high-volume, lower-stakes content. A reasoning model for complex or borderline cases where accuracy matters more than latency. Most teams use a mix, and some escalate from one to the other.
• You make the enforcement decisions. We connect via a single API: you send content, we send back the outputs (label, rationale, severity, confidence). You decide what to do with them. Auto-enforce on the clear cases, route the mid-confidence ones to human review.
• It includes tooling to help you write better policies. We've built a policy optimizer for the finicky work of turning community guidelines into something an LLM enforces predictably.

Our tooling is a good fit if you're dealing with policy complexity, nuance that simple classifiers can't handle, or a need to explain decisions to users, regulators, or your own team.

Fraud & Spam Enforcement: for getting ahead of bad actors

If your main challenge is fraud, spam, scam accounts, bots, or fake profiles, and you feel like you're always one step behind, our fraud layer is built for you.

Instead of enforcing written rules, it learns. It watches what your moderation team does, combines that with behavioral signals (how fast is this user messaging? how long after signup did they start?) and metadata patterns (email characteristics, IP patterns, device signals), and builds a model of what bad looks like in your specific community.

Because it learns from your moderators rather than a generic training set, it adapts to the particular tactics people use to abuse your platform, which tend to be specific to your vertical and user base. Each decision comes back as a risk score and the underlying signals behind it, so you can set thresholds and route accordingly. We continually look at agreement levels between our model and your moderators, to ensure that there's alignment.

You can use it proactively (score accounts at signup, at first message, at key behavioral triggers) or reactively (after a report comes in). Most teams do both.

Real-Time Chat Moderation: for ultra-high-volume live surfaces

If you're moderating live chat (like in-game player-to-player, livestream chat when a moment goes big, live support sessions, comment streams, usernames and profiles at signup) you've probably hit the wall where the usual options don't fit. LLM moderation is steerable but too slow and costly at billions of messages a day. Classic classifiers are fast and cheap but locked to a fixed taxonomy, so changing what counts as a violation means retraining and relying on a vendor or engineering team to pay attention to your ticket.

This is the third option: the steerability and control of an LLM at the cost and latency of a classifier, with median decisions under 150ms on dedicated deploys. That's the difference between blocking or holding content inline, before it renders, and pulling it down after users have already seen it.

You get two ways to adapt. Edit your policy in plain language and the model picks up the change instantly ("harassment, but not in-game trash talk" becomes a policy edit, not an engineering ticket). Or retrain from your team's own relabeling when you want a deeper shift. On top of that: a configurable 0–10 risk score to route on, per-category thresholds, the freedom to add or remove categories without starting over, and multilingual coverage out of the box.

It shares policies with our content moderation, too. Author once, route high-volume real-time traffic to the fast model, and send gray-area calls up to the LLM tier.

Content Insights & Analytics: for seeing what's actually there

The content flowing through your platform is rarely what you picture. You know the content that fits the categories you already track, but that's a fraction of what's really there.

It plots all your content into a cluster map, grouping it by what it actually means rather than by format, label, or language. The same harm in five languages or three formats lands as one cluster instead of scattering into noise. Zoom into any cluster to see what's inside and jump to the decisions and policies behind it. When a cluster spikes with a new scam or coordinated campaign you've never named, it flags what you didn't know to look for, before it spreads.

• You don't need to be a Musubi moderation customer to use it. It's the one part of Musubi where you point us at your data to understand it rather than to label it, so it layers over whatever pipeline you already run.
• It's not a BI dashboard. BI tools count the categories you already defined. This shows you what you haven't named yet.
• It's a shared lens, not just an enforcement tool. Safety teams catch emerging harm and QA their moderation; product, analytics, and leadership get an honest read on what users are really doing. Data is ingested in real time, pinned views auto-refresh, and you can export any cluster to CSV.

Good fit if you want visibility ahead of a known risk window (an election, launch, or major live event) or you suspect your current categories are missing things.

Agentic Investigations: for cases that take more than one step

Some cases can't be settled with a single fast call on a single piece of content. They need several steps and outside context: visiting a linked page, reading a document, pulling a record from your systems, and reasoning across all of it before recommending an action. Think ad review, document and ID forensics, or untangling a coordinated abuse ring.

Give the agent a single signal and it works the case across steps, cross-references your data, and assembles the full case into one recommendation with a reasoning trace. It replaces a lot of the grunt work of investigations, and leaves your team to make the final call.

• It's steerable to your platform. You set the policies it enforces, the signals it weighs, the actions it can take, how aggressive it is, and where it must stop for human review.
• It doesn't get to invent evidence. Findings are grounded in actual tool use, with citations, and every investigation comes with a full reasoning trace a human can verify.
• It can use your own data sources. Agents can query your internal systems and records mid-investigation, so a case is judged against what you actually know, not just what's publicly visible.
• Common deployments: document & ID review, ad review, coordinated abuse, account takeover, or a custom agent for any case type your team faces.

Moderation Console: for teams starting without one

If you don't already have a moderation console with review queues, dashboards, and orchestration for your team to actually work in, this is the full system.

It's a hosted version of an open-source moderation platform built and maintained by ROOST and the T&S community. You get the transparency of open source (see exactly how it works, shape it to your platform, no black box) without standing up or maintaining any of the infrastructure yourself. We host it, keep it patched, and support you directly.

What's in the box: review queues with role-based routing, a rules engine you build in the UI, dashboards and analytics, DSA transparency reporting and appeals, and reviewer wellbeing controls (media blurred by default, plus grayscale and auto-mute you can set org-wide or per reviewer). Child-safety tooling is built in too, with HMA hash matching to catch known CSAM without anyone having to view it, plus Google and OpenAI integrations and CyberTipline reporting to NCMEC.

Most teams start with the console and switch on Musubi's AI moderation and fraud detection later, as decision signals, with no re-integration. The console stands on its own, but the advantage of getting it from us is that the rest of Musubi plugs straight in when you're ready.

Workflows: for tying it all into one process

As you add more of these tools, you need a way to connect them into a process that matches how your team actually works. This is the workflow engine that does it.

It lets you build, automate, and scale Trust & Safety processes across tools, teams, and use cases. Route enforcement through different models depending on the content or the account, escalate to human review at the points where judgment matters, and chain the steps together into a flow you can reshape as your policies change. Instead of hard-coding your process into engineering tickets, you build it once and adjust it as you go.

Good fit if you're running more than one signal or tool and want them orchestrated into a single, adjustable pipeline rather than a set of disconnected checks.

Using them together

The pieces are more powerful together, and they can talk to each other.

The most common pattern: content moderation is removing a user's content over and over, but the account itself hasn't crossed any threshold. That pattern is a signal worth escalating, so the account gets sent to fraud enforcement for a fuller review. If it agrees, you take an account-level action instead of playing whack-a-mole with individual posts.

The reverse happens too. An account looks borderline to fraud enforcement but not clearly enough to act on, so the system pulls the content labels for that account. Has this user been posting things that also look problematic? The combined picture is usually much clearer.

From there it stacks: content insights surface the emerging pattern before it's named, agentic investigations work the hardest multi-step cases, and the moderation console gives your team one place to review and act on all of it. If you're dealing with adversarial users who know how to stay just below the line on any single signal, running these in a loop closes a lot of the gaps.

Integrations

We're built to plug into the console your team already uses, whether that's your own in-house tooling or a partner platform. We integrate with Tremau's Nima platform, which is a great choice, as well as NVIDIA's NeMo guardrails. And if you don't have a console yet, our moderation console (above) is the answer for starting from scratch.

A note on what we're not

We're not a pre-built classifier you buy off the shelf. If you want to point an API at your content and get preset labels back, there are other options for that.

What we do is give your team real control over the policies, the logic, and what the AI learns from. That takes some upfront work and configuration. In return, you get a system that reflects your platform's values and adapts to your specific threat environment, instead of a generic model you're constantly fighting.