Skip to content

Media & Streaming

Moderation agent

A moderation agent that triages the queue automatically and routes only the genuinely ambiguous cases to human reviewers — turning a backlog that grew with every upload spike into a steady flow.

  • LangGraph
  • OpenAI
  • River
-91%

time-to-decision

reviewer throughput

6×

In short

  • Unambiguous cases are decided automatically; reviewers see only the genuinely borderline.
  • The queue is ranked by impact, so the decisions that matter are handled first.
  • Durable processing absorbs upload spikes without dropping items.

The engagement

Sector
Media & streaming (UGC)
Company shape
Platform with spiky upload volume
Engagement
Triage agent + reviewer tooling

Anonymised by policy

Every engagement runs under NDA — we protect our clients' confidentiality as a matter of policy, and we'd extend the same discretion to your work. The constraints, the shape, and the outcome here are real; the identity is redacted.

The challenge

Every upload spike grew the moderation backlog faster than reviewers could clear it, and the delay hit exactly the content that mattered most. Blanket automation risked wrong calls on edge cases; pure human review couldn't scale.

The approach

  1. 1
    Automate the clear cases

    The agent decides the unambiguous cases directly and reserves human attention for the genuinely borderline ones.

  2. 2
    Prioritise the queue

    Ambiguous items are ranked so reviewers see the highest-impact decisions first, not just the oldest.

  3. 3
    Absorb spikes

    Durable job processing (River) lets the system soak up upload spikes without dropping items or falling over.

The solution

A LangGraph agent classifies each item, auto-decides the clear cases, and routes the ambiguous ones to a prioritised reviewer queue. Durable processing absorbs spikes, and every decision is traceable for audit.

The system we built

  1. Ingestupload event
  2. Classifypolicy check
  3. Auto-decideclear cases
  4. Prioritised reviewambiguous only
Abstract by design — the architecture we built, not a client screen. No confidential data is shown.
Illustrative
Triage in action: auto-decide the clear cases, escalate the rest
Illustrative
The prioritised reviewer queue — highest-impact decisions first

The results

time-to-decision

-91%

median, vs. manual triage

reviewer throughput

6×

decisions per reviewer-hour

Data & safety

Guardrails
The agent auto-decides only unambiguous cases; borderline items route to a prioritised human queue.
Evaluation
Policy checks are versioned and every decision is traceable for audit.
Data handling
Durable job processing soaks up spikes; no item is lost to a timeout.

The stack

  • LangGraph
  • OpenAI
  • River

Questions we get

The agent decides only the unambiguous cases and routes anything borderline to a prioritised human queue — automation is bounded to what it can judge reliably.

Ambiguous items are ranked so reviewers see the highest-impact decisions first, not just the oldest — the backlog stops being FIFO.

Durable job processing absorbs spikes without dropping items, and every decision is traceable for audit after the fact.

More selected work

// Let's build

Want an outcome like this on your problem?

Tell us where you are. We reply within a day with a concrete next step.