What this system is actually solving
Pure collaborative filtering struggles with new users and new content. Pure content similarity ignores behavioral preference. This project combines both while preserving diversity.
The project is designed around an observable outcome and explicit constraints. A production decision is only considered successful when the model or algorithm improves a baseline without creating unacceptable cost, latency, reliability or interpretability problems.
How the system is decomposed
- 01Event collector
A separately testable stage with defined inputs, outputs, observability and failure handling. - 02Content feature index
A separately testable stage with defined inputs, outputs, observability and failure handling. - 03User interest profile
A separately testable stage with defined inputs, outputs, observability and failure handling. - 04Candidate generators
A separately testable stage with defined inputs, outputs, observability and failure handling. - 05Learning-gap scorer
A separately testable stage with defined inputs, outputs, observability and failure handling. - 06Diversification layer
A separately testable stage with defined inputs, outputs, observability and failure handling. - 07Ranking service
A separately testable stage with defined inputs, outputs, observability and failure handling.
Why each algorithm exists
TF-IDF/content embeddings
Role. Places related concepts near one another in vector space so retrieval can capture semantic similarity.
Trade-off. Embedding quality, domain mismatch and vector-index configuration strongly affect results.
Evaluate with. Recall@k, nDCG, semantic benchmark accuracy and latency.
Cosine similarity
Role. Compares vector direction rather than raw magnitude, making it useful for normalized text representations.
Trade-off. Its quality is limited by the representation placed into the vector space.
Evaluate with. Pairwise relevance accuracy and retrieval ranking quality.
Weighted implicit feedback
Role. Weighted implicit feedback is included because it addresses a specific measurable part of the system rather than being added as decoration.
Trade-off. The component must be compared with a simpler baseline and removed if it adds complexity without measurable value.
Evaluate with. Task-specific quality metric, latency, reliability and failure-case analysis.
MMR-style diversification
Role. MMR-style diversification is included because it addresses a specific measurable part of the system rather than being added as decoration.
Trade-off. The component must be compared with a simpler baseline and removed if it adds complexity without measurable value.
Evaluate with. Task-specific quality metric, latency, reliability and failure-case analysis.
Contextual bandit exploration
Role. Uses feedback to allocate more traffic to promising choices while reserving some exploration for alternatives.
Trade-off. Biased or sparse rewards can push the policy toward a locally attractive but globally poor choice.
Evaluate with. Reward lift, regret, exploration coverage and stability over time.
Signals entering the system
Every useful model depends on the quality, timing and provenance of its inputs. CortexLab treats feature definitions and leakage checks as part of model engineering, not preprocessing trivia.
- Reads — captured with validation, lineage and monitoring so training and production meaning stay aligned.
- Bookmarks — captured with validation, lineage and monitoring so training and production meaning stay aligned.
- Interview weaknesses — captured with validation, lineage and monitoring so training and production meaning stay aligned.
- Game performance — captured with validation, lineage and monitoring so training and production meaning stay aligned.
- Topic metadata — captured with validation, lineage and monitoring so training and production meaning stay aligned.
- Completion events — captured with validation, lineage and monitoring so training and production meaning stay aligned.
What has to be measured before calling it successful
- CTR
- Completion rate
- Coverage
- Diversity
- Novelty
- Long-term learning gain
Evaluation is split between offline quality, operational performance and failure analysis. A high headline metric does not override poor calibration, unstable segments, leakage or unusable latency.
Where this project can fail
Cold start needs content signals
This risk is tracked through tests, monitoring, explicit thresholds or human review depending on where it appears in the architecture.
Engagement is not the same as learning value
This risk is tracked through tests, monitoring, explicit thresholds or human review depending on where it appears in the architecture.
Ranking needs diversity constraints
This risk is tracked through tests, monitoring, explicit thresholds or human review depending on where it appears in the architecture.
How the project grows without becoming untestable
- 01Sequence-aware ranking
- 02Counterfactual evaluation
- 03Bandit personalization
- 04Graph recommendations
Questions an engineer should be able to answer
- Why is this architecture preferable to a simpler baseline?
- Which metric can look good while the product still fails?
- Where can data leakage enter this pipeline?
- What changes when traffic, data volume or latency requirements increase 10×?
- Which part should be rolled back first if production quality drops?