Abstract real-time data streams
Two-stage retrieval · Real-time feature pipelines

Recommendations at
the speed of thought.

StreamMind Ultra is a Netflix-grade recommendation engine. It blends a two-tower retriever, a gradient-boosted ranker, and real-time Kafka pipelines to serve personalized picks in milliseconds — with full MLOps under the hood.

Sub-millisecond ANN search Kafka streaming MLflow + Airflow + DVC
2-stage
Retrieve → Rank pipeline
<1ms
FAISS ANN lookup
Top-20
Ranked from 500 candidates
Real-time
Kafka feature updates
The platform

A recommender built like the big leagues.

Two-stage recommender

A Two-Tower + Matrix Factorization retriever generates candidates; a LightGBM / deep ranker scores P(click) to return the top 20.

Real-time personalization

Apache Kafka streams every watch, click, and rating into the feature store — updating recommendations in milliseconds.

Vector similarity search

User and item embeddings live in a FAISS index for sub-millisecond approximate nearest-neighbor retrieval at scale.

Online + offline feature store

Feast serves low-latency features from Redis while versioning batch-computed training data in PostgreSQL and Parquet.

A/B testing engine

Run champion/challenger experiments with deterministic traffic splits and live click-through-rate tracking per variant.

Full MLOps stack

MLflow experiment tracking, DVC data versioning, and Airflow DAGs orchestrate retraining, embeddings, and deployment.

StreamMind · Serving dashboardlive · 12.4k req/s
Recommended for user u_10427
Neon Horizon
Sci-Fi98%
The Quiet Hour
Drama94%
Deep Current
Thriller91%
Paper Cities
Indie88%
CTR (24h)
18.7%
+3.2% vs control
P99 latency
42ms
retrieve + rank
The architecture

From raw event to ranked pick in four hops.

01

Ingest events

Watches, clicks, ratings and searches stream into Kafka topics in real time from the API and dashboard.

02

Update features

Consumers compute fresh features into the Redis online store while Feast versions batch features offline.

03

Retrieve candidates

The Two-Tower model embeds the user; FAISS returns the top 500 nearest items in under a millisecond.

04

Rank & serve

A LightGBM ranker scores each candidate on affinity, recency and popularity, returning the top 20 picks.

Real-time monitoring visualization
Observe & scale

Production instrumentation, not a demo notebook.

StreamMind ships with the operational muscle real systems need: end-to-end metrics, drift alerts, and a deployment workflow that lets you roll models forward — and back — with confidence.

  • Prometheus metrics for latency, CTR, and cache hit-rate
  • Grafana dashboards with model and feature drift detection
  • Microservices you can scale independently on Kubernetes
  • Champion/challenger deployment with automatic rollback

Spin up the whole stack in one command.

Kafka, Redis, PostgreSQL, the API, MLflow, Airflow and Grafana — orchestrated with Docker Compose. Open source under MIT.

bash
git clone https://github.com/shahriar-ahmed-seam/streammindcd streammind && docker-compose up -d