Key Takeaways

  • Real reaches 1.5 million monthly active users by focusing on fantasy players, sports bettors, and die-hard fans who want live score data paired with fast social interaction.
  • While typical social apps struggle to get 1% of their audience to post, Real converts 30% of its monthly user base into active commenters.
  • Legacy sports apps like ESPN and Bleacher Report limit push notifications to halftime or final scores; Real triggers alerts during live micro-streaks, such as a player scoring 12 consecutive points.
  • Instead of selling standard static player cards, the platform monetizes through a gaming model inspired by Roblox and Fortnite, letting users buy and collect specific digital moments like a single touchdown or three-pointer.
  • To protect chat speed and fan banter during games, the founders deploy Real's Multi-Tiered Real-Time Moderation Architecture.

The Real Multi-Tiered Real-Time Moderation Architecture

Co-founder Louis Antonelli points out that legacy sports apps treat live scores as passive reference sheets. Real treats games as synchronous social events. When thousands of fans react to a single play simultaneously, standard API calls to large language models create latency bottlenecks that destroy the live experience. To solve this, Real routes incoming comments through three distinct filters:

  • Tier 1: Lookup Tables: A high-speed lookup table system that automatically catches and filters 80% to 85% of baseline profanity and toxic terms in real time.
  • Tier 2: Fast Classifier: A lightweight secondary classification model that evaluates context and catches an additional 10% of disruptive comments.
  • Tier 3: Frontier Model Escalation: Any borderline or ambiguous comments that pass through the first two layers are escalated to modern foundation/LLM models for final moderation decisions.

“A lot of the ESPNs, Bleacher Reports, their score apps would never send real-time notifications, only halftime or end of game moments,” Louis Antonelli noted. By pairing sub-second play alerts with low-latency chat moderation, the app keeps fans inside the stream during the game rather than checking a box score after the final whistle.

When This Works (and When It Doesn't)

This architecture works in high-velocity live chat environments with an 18-to-30 demographic, especially in sports, gaming, and live events. In these spaces, latency is fatal. If a fan's message takes three seconds to clear an AI safety check, the game has already moved on and the conversation is dead. The tiered model filters obvious spam instantly while preserving competitive banter, trash talk, and fan culture.

This method fails in asynchronous or low-volume environments. If you run a B2B community, a professional network, or a long-form discussion forum, latency does not matter. Routing everything through Tier 1 lookup tables in a professional setting creates false positives on industry jargon while missing subtle harassment. In those products, skip the lookup tables and send every comment directly to a frontier model with your full community guidelines in the system prompt.

What to Do With This

If you are building a real-time social product or live feature this week, audit your moderation latency:

1. Measure your current time-to-publish for user comments. If an LLM call adds more than 200 milliseconds to message delivery, your chat will feel sluggish during traffic spikes.

2. Build a hardcoded Tier 1 blocklist for unambiguous slurs and spam patterns. Set this filter in memory so it executes in single-digit milliseconds.

3. Train a small, local classification model for Tier 2 to handle slang and sarcasm specific to your niche.

4. Route only the remaining 5% to 10% of ambiguous edge cases to a paid LLM API. This keeps latency near zero and cuts your AI inference costs by up to 90%.