Sources: Netflix Investor Relations · Zhenzhong Xu, Netflix Engineering · Netflix Research
Netflix didn't just build a streaming service. They built a planetary-scale similarity-routing engine. Every time you open Netflix, a massive distributed system computes which content you're most likely to enjoy by comparing you to users with similar viewing patterns.
The technical achievement is staggering: 2 trillion events per day processed through the Keystone data pipeline, collaborative filtering across 250 million user profiles, and personalized recommendations served in under 100 milliseconds. 75-80% of all viewing hours come from algorithmic recommendations.
QIS uses the same fundamental infrastructure patterns—similarity-based routing at scale—but for a different purpose: routing insight instead of entertainment. The architecture is nearly identical. The paradigm is fundamentally different.
The Full Netflix Loop: End-to-End Architecture
Key technologies: Apache Kafka (ingestion) → Apache Flink (stream processing) → Apache Spark (batch training) → Matrix factorization + K-NN (similarity) → Microservices (serving) → Open Connect CDN (delivery). The K-NN step finds the K nearest neighbors in latent vector space—intentionally fuzzy matching for serendipitous discovery.
The Full QIS Loop: End-to-End Architecture
Key insight: Any data source → Edge node (phone, hospital, vehicle) → Expert-defined fingerprint → Any similarity router (DHT, vector DB, etc.) → Outcome packets (~512 bytes, the insight itself) → Local synthesis. Your situation becomes your address. The parameters that define your problem deterministically route you to everyone who shares that problem. Same address = same cohort = relevant outcomes waiting for you.
The Paradigm Shift: Same Infrastructure, Different Payload
What Gets Routed Changes Everything
🎬 Netflix Routes
Content IDs
"Movies you might like based on users similar to you"
Recommendations
🔮 QIS Routes
Outcome Packets
"What actually happened to people exactly like you"
Insight Itself
Netflix's payload is a content ID—a pointer to a movie in their catalog. The recommendation says "you might like this." The user still has to watch the movie to find out.
QIS's payload is the outcome itself—a pre-distilled insight that already contains the answer. "Treatment A worked. 18 months progression-free. Confidence 0.94." The insight doesn't point to something you need to go find. The insight IS what you came for.
The Routing Infrastructure Doesn't Compute
In both systems, the routing layer is a delivery mechanism—not a processor. Netflix's K-NN doesn't watch movies for you. QIS's DHT doesn't analyze outcomes. They route. The difference is what arrives when routing completes. Netflix delivers suggestions. QIS delivers answers.
The Precision Distinction: K-NN vs Exact Cohort
This is where the architectures diverge by design choice, not capability.
🎬 Netflix: K-Nearest Neighbors (Fuzzy)
- Method: Find the K closest users in latent vector space (cosine similarity)
- Typical K: 10-50 neighbors
- Intent: Discovery and serendipity—"you might also like"
- Precision: Intentionally approximate
- Why: A user who likes Stranger Things should discover The OA even if they're not identical matches
- Stakes: Wrong recommendation = mildly annoying. No harm.
🔮 QIS: Exact Cohort Matching (Precision)
- Method: Deterministic hash of categorical variables → exact bucket routing
- Bucket size: All peers with identical fingerprint (could be thousands)
- Intent: Precision matching—"outcomes from people exactly like you"
- Precision: Stage 3 NEVER matches Stage 4
- Why: Medical safety requires categorical boundaries. KRAS+ must not mix with KRAS-.
- Stakes: Wrong cohort = potentially dangerous. Precision is non-negotiable.
Both systems CAN do either. Netflix could do exact matching—but chose not to because entertainment discovery benefits from serendipity. QIS can do approximate matching (vector DB path)—but emphasizes exact routing because healthcare and safety-critical domains demand precision.
The "Teleportation" Principle
When your fingerprint is deterministic and your routing preserves its structure, you don't search the network. You don't wander. You teleport directly to where relevant outcomes already exist. The expert defines what "similar" means. The hash makes it exact. The DHT routes you there in O(log N) hops. Everyone in that bucket shares your expert-defined situation. Their outcomes are waiting for you.
Side-by-Side Technical Comparison
| Component | Netflix | QIS |
|---|---|---|
| Data Ingestion | Apache Kafka (700B+ events/day, 36 clusters) | Any source → Edge node (local aggregation) |
| Stream Processing | Apache Flink (real-time ETL) | Edge node (local distillation) |
| Batch Processing | Apache Spark (model training) | None required (outcomes pre-distilled) |
| Similarity Definition | Matrix factorization → latent vectors | Expert template → deterministic address |
| Matching Method | K-NN in vector space (fuzzy, K=10-50) | Exact bucket + continuous refinement |
| Routing | Microservices + Eureka discovery | DHT, Vector DB, or any similarity router |
| Routing Complexity | O(1) cached + O(K) re-ranking | O(log N) hops to exact bucket |
| Payload | Content IDs (recommendations) | Outcome packets (the insight itself) |
| Serving Latency | <100ms | Similar (routing + synthesis) |
| Local Computation | Display logic | Outcome synthesis (vote, aggregate) |
| Central Model | Required (trained on full dataset) | None required (outcomes pre-exist) |
| Privacy | Data to Netflix for training | Raw data stays local (only outcomes route) |
Universal Applicability: Any Domain Where Similarity Is Definable
Netflix proved the infrastructure works at planetary scale for entertainment. The same architectural pattern applies to any domain where you can define similarity and insight is aggregatable.
🏥 Precision Medicine
Route to patients with identical clinical fingerprint (stage, biomarkers, mutations). Retrieve real treatment outcomes. "What worked for people exactly like me?"
🌾 Agriculture
Route to farms with identical conditions (soil type, climate, crop variety). Retrieve yield outcomes. "What intervention worked for fields exactly like mine?"
🔧 Industrial Machinery
Route to equipment with identical profile (model, age, usage pattern). Retrieve failure/maintenance outcomes. "What preventive action worked for machines exactly like mine?"
🏛️ Prison System
Route to inmates with identical profile (offense type, risk factors, history). Retrieve intervention outcomes. "What rehabilitation program worked for people exactly like this?"
🚗 Autonomous Vehicles
Route to vehicles that encountered identical scenarios (road conditions, edge cases). Retrieve outcomes. "What action succeeded in situations exactly like this?"
📊 Financial Risk
Route to portfolios with identical exposure profiles. Retrieve historical outcomes. "What happened to positions exactly like mine in similar conditions?"
✈️ Aviation
Route to flights with identical parameters (aircraft type, weather, route, payload). Every plane before takeoff already knows how every similar plane landed. Real-time turbulence outcomes, fuel efficiency patterns, approach success rates—across every airline, globally.
🎓 Education
Route to learners with identical profiles (background, learning style, struggle patterns). Every student stuck on a concept instantly sees what worked for every similar learner. Not generic advice—specific interventions that produced breakthroughs for people exactly like them.
From Coughs to Crops to Cars
The infrastructure is domain-agnostic. If you can define similarity (what makes two cases "the same") and insight is aggregatable (outcomes can be synthesized), QIS applies. Netflix built a planetary-scale proof of concept for entertainment. The same pattern extends to any domain where distributed data holds distributed insight.
The Bottom Line
Netflix didn't just build a streaming service. They built infrastructure for similarity-based routing at planetary scale—2 trillion events per day, 250 million users, sub-100ms latency. They proved the architecture works.
QIS is the protocol for scaling intelligence itself.
Not an application. Not a product. A universal mechanism for turning distributed experience into collective insight—anywhere similarity can be defined and outcomes are valuable. Netflix routes entertainment. QIS routes what humanity has learned. The infrastructure is identical. The implications are not.
Netflix routes content recommendations using fuzzy K-NN matching because serendipity serves entertainment.
QIS routes outcome packets using exact cohort matching because precision saves lives.
Same infrastructure. Different paradigm. The technology exists. The proof is running at scale. Netflix asked: "What might you like?" QIS asks: "What actually works for people exactly like you?"
From coughs to crops to cars to classrooms to cockpits—Netflix proved the infrastructure works at planetary scale. QIS points it at what matters.
Imagine a world where insight flows freely and intelligence scales quadratically—in real time, for everyone, everywhere.
Imagine That World →