Just shipped a real-time intent classifier that adapts to what you're reading — not just what you've liked in the past. The signal decay model is the key insight. The further back an interaction is, the less it should define your current feed. Open-sourcing the decay function next week. 🧠
Why most social feeds fail at discovery: → They optimize for what you already know you like → Not what you're about to become interested in → The gap between "liked before" and "exploring now" is where all the interesting discovery happens NEXUS's IntentGraph is the first real attempt I've seen to close that gap.
The difference between a social network and an intelligent social network is whether it can tell the difference between who you were and who you're becoming. Most platforms answer the question "what did this person engage with?" NEXUS is trying to answer "what does this person need to find right now?" That's a fundamentally different product.
"The platform learns permitted signals about what a person is interested in now." — NEXUS Blueprint
Graph databases for social networks: when your social graph and intent graph share the same traversal primitives, discovery gets genuinely interesting. The query "find me content that my follows' follows are engaging with, filtered by my current intent vector" is 3 hops in a graph DB. In a relational DB it's 4 joins and a subquery. Architecture matters for discovery.
MATCH (me:User {id: $userId})-[:FOLLOWS*1..2]->(creator:User)
<-[:AUTHORED]-(post:Post)-[:TAGGED]->(topic:Topic)
WHERE topic.id IN $intentTopics
AND post.created_at > $cutoff
RETURN post ORDER BY post.score DESC LIMIT 50Hot take: the reason social media feels exhausting is not the content — it's the mismatch between what the algorithm thinks you want and what you actually need right now. When your feed knows the difference between "I always liked cooking videos" and "I'm specifically looking for quick weeknight recipes this week" — that's when it stops feeling like work.
Running NEXUS's recommendation scoring at scale notes: • Intent match component is the heaviest compute — embedding similarity is O(n) against user intent vector • Caching the intent vector per-user with a 5-minute TTL drops p99 latency from 340ms to 94ms • Redis sorted sets for the candidate pool are the right call — O(log n) insertion and range queries The 94ms median isn't magic. It's cache design.
Trust architecture in social platforms is not a feature — it's the foundation. NEXUS's Trust Engine is designed in from day one: reports, moderation queues, audit logs, appeals paths. The cost of retrofitting this after launch is 10× the cost of building it correctly upfront. The platforms that got this wrong are still paying that debt.
The NEXUS business model is interesting: they're not selling attention to advertisers as the primary loop. They're selling useful discovery. When a business shows up in your feed because your current intent matches what they offer — not because they paid for placement — that's a fundamentally different value proposition. Qualified leads > raw impressions.
Security note for anyone building on Supabase at scale: Row Level Security is not optional — it's the boundary between your users' data. Test both the allowed AND the denied cases. The Supabase security advisor will catch the obvious gaps, but edge cases in complex join policies need manual review. Founder control principle: the service-role key never touches the browser. Ever.
First month on NEXUS as a creator: • 3.2K followers from 0 • 47 posts published • Average IntentMatch score: 84% • Best performing topic: #distributed-systems The difference from other platforms: I can actually see which topics are driving discovery vs. which ones my existing followers engage with. That distinction is invaluable for content strategy.
Your current interest distribution — updated as you engage.