Stars
Forks
Watchers
Developer links
OpenAdServer
OpenAdServer runs your own ad network on your own infrastructure, ranking every request with machine-learned click predictions instead of static rotation. A request arrives with a slot, user, device, and geo context; the pipeline retrieves matching campaigns, filters them by targeting, budget, and frequency caps, predicts click probability, and ranks survivors by eCPM before responding in milliseconds. You serve banner, native, video, and interstitial formats with geo, device, operating system, demographic, and interest targeting, plus daily and hourly frequency limits per user. Budget pacing smooths delivery across the day so campaigns do not exhaust spend in the first hour. Three prediction models ship ready to train and swap without downtime: logistic regression for speed, factorization machines for feature interactions, and DeepFM for deep learning on your click logs. Automatic feature engineering processes sparse and dense fields, and hot-swapping lets a retrained model replace the live one mid-flight. Impressions, clicks, and conversions stream into event tracking, Prometheus metrics expose full observability, and Grafana dashboards visualize performance in real time. Advertisers, campaigns, and creatives are managed through CRUD APIs, and bid types cover CPM, CPC, CPA, and optimized CPM. Training scripts prepare data from your logs, train a model in a few commands, and report AUC before you promote it. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.
Benefits
- Revenue ranked by machine learning
- Every request is scored by a CTR model and ranked by eCPM across CPM, CPC, and CPA bids, so higher-value impressions win without manual yield tuning.
- Keep all of the revenue
- Self-hosting removes the network revenue share entirely; your campaigns, click logs, and models stay on infrastructure you control with no platform percentage taken.
- Latency built for the page
- The retrieval, filter, predict, and rank pipeline answers in single-digit milliseconds at the prediction step, keeping ad calls from slowing the pages that carry them.
- Retrain and swap live
- Training scripts turn your event logs into a fresh model, evaluation reports AUC first, and hot-swap promotes it without restarting the serving process at all.
Features
- Ad serving pipeline
- Retrieve, filter, predict, rank, and rerank stages process each request with targeting, budget pacing, and frequency capping applied in order.
- CTR prediction models
- DeepFM, factorization machine, and logistic regression models with automatic sparse and dense feature engineering and sub-5ms inference.
- Smart targeting
- Geo, device, operating system, demographics, and interests narrow delivery, with daily and hourly frequency caps per user.
- Event tracking
- Impression, click, and conversion events are recorded per request and ad, feeding analytics, model training data, and billing reports.
- Campaign management APIs
- CRUD endpoints for advertisers, campaigns, and creatives, plus health and Prometheus metrics endpoints for operations.
- Model training toolkit
- Scripts prepare training data from logs, train and evaluate models with AUC and log-loss reporting, and benchmark throughput per model.
