SafePulse — Real-Time Community Safety

Inspiration

Online communities move fast—but community safety systems don’t.

During moments of high activity—product launches, outages, breaking news, or emotional discussions—harmful content can escalate within seconds. Toxicity, harassment, and distress signals often spread long before moderators or batch analytics tools notice a problem.

We were inspired by a simple realization: community safety is a real-time problem, but most tools are built for after-the-fact analysis. We wanted to explore what happens when AI operates continuously on live conversations instead of static datasets.

That question became the foundation for SafePulse.

What it does

SafePulse is a real-time community safety platform that detects and responds to harmful behavior as it happens.

It continuously monitors live community conversations (such as public Reddit posts and comments) and:

• Classifies toxicity, harassment, hate speech, and potential self-harm risk
• Tracks real-time safety trends across communities
• Detects sudden spikes in harmful behavior
• Triggers immediate alerts and moderation signals

Instead of waiting for reports or delayed dashboards, SafePulse provides instant visibility into community health, enabling faster and more effective responses.

How we built it

SafePulse is built around the idea of AI on data in motion.

Real-time ingestion

Public community conversations are ingested as a continuous stream, where every post or comment is treated as an event.

Event streaming backbone

We use Confluent Cloud as the real-time backbone to ingest, buffer, and distribute events reliably at scale.

Stream processing

Using stream processing (Flink / ksqlDB), we clean, enrich, and aggregate events over rolling time windows to detect patterns such as toxicity spikes or rising risk levels.

AI enrichment

Each event is enriched with safety signals using Google Cloud Vertex AI, which classifies:

• Toxic or abusive language
• Harassment and hate speech
• Potential self-harm indicators

Confidence scores help prioritize the most urgent cases.

Real-time responses

When predefined thresholds are crossed, SafePulse emits alerts and safety actions that can notify moderators, trigger workflows, or update live dashboards.

This architecture allows AI to operate continuously, not in batches.

Challenges we ran into

Near real-time data ingestion

Public platforms like Reddit don’t provide true push-based streaming APIs, so we had to design a near-real-time polling approach that respected rate limits while still feeling live.

Balancing speed and accuracy

Applying AI models in real time required careful trade-offs to keep latency low

Built With

• ai
• apache-kafka
• community-safety
• confluent-cloud
• confluent-cloud-(kafka-+-flink)
• google-cloud
• python
• real-time-streaming
• vertex-ai