Contact person: Business Client

Phone:Show

Email:Show

Location: United States

Budget: Recommended by industry experts

Time to start: As soon as possible

Project description:
"Summary
Recommendation and ranking project.

Focus: build the ranking and data backbone that makes the algorithm feel “alive” — real-time features, logistic-regression ranking, and nightly retraining based on user interactions.


Tech Stack

Infra: Nats, Redis, Postgres or MongoDB, S3/GCS

Compute: Python, PyTorch / scikit-learn, Flink / Nats

Data: Parquet / Delta, Metabase, Feature Store (Feast-style Redis + S3)

Model Serving: ONNX / FastAPI / Triton

CI/CD: ArgoCD, GitHub Actions



You’re a Fit If You:

Have built production recommendation or ranking systems (newsfeed, ads, marketplace, video, etc.).

Know streaming data systems (Kafka, Flink, or Spark Structured Streaming).

Can implement online metrics (EMA, sketches, windowed aggregates).

Understand point-in-time feature joins and data leakage prevention.

Have shipped simple but high-impact models (logistic regression, tree-based, or two-tower retrieval).

Are hands-on — able to go from raw events to a deployed model in a few days.

Are pragmatic: prefer a working baseline this week over a perfect pipeline next month.

Deliverables
Implement the data → features → model → inference loop end-to-end.
Stand up real-time stream feature computation (Kafka, Flink, or Kafka Streams).
Build rolling metrics (CTR, finish rate, popularity EMAs) into Redis for live ranking.
Train and deploy a logistic regression model using nightly batches from S3/Parquet.
Implement isotonic calibration for well-behaved prediction scores.
Create a reproducible training pipeline with point-in-time joins and schema versioning.
Integrate model inference into the Feed API (Typescript microservice).
Monitor feature freshness, Nats lag, and model staleness SLIs." (client-provided description)