Skip to content

About

course-recommender-lab is a content-based recommender-systems study on the UC Berkeley course catalog, built by Sandeep Jayaprakash as part of a portfolio of production-pattern ML and data-engineering projects.

Why a "lab"

The name is deliberate. This isn't a single model shipped once — it's a controlled environment where eight technique families are held to one interface, scored by one evaluation harness, and ranked on one leaderboard. The value is in the comparison and the methodology: getting the measuring stick right, enforcing leakage discipline, reporting confidence intervals, and treating negative results as first-class findings rather than hiding them.

Design principles

  • Honest documentation. Limitations and failures are first-class. Three techniques that should have won and didn't are the most prominent results, not footnotes.
  • Production patterns from day one. A swappable ABC, persisted artifacts, pinned dependencies, an ADR per architectural decision, tests alongside every technique.
  • Reproducible and key-free. Global RANDOM_SEED = 42; the whole repo runs end-to-end with no API key. The LLM technique uses a local Ollama daemon and degrades gracefully when it's absent.
  • Portable. No hardcoded paths, keys, or magic numbers; everything runs on a laptop.

The rest of the portfolio

  • scribe-iq — a grounded clinical documentation AI prototype: synthetic Synthea patient spine, public clinical-note corpora, RAG over pgvector, FastAPI + Next.js, multi-cloud LLM providers, and governed audit workflows.
  • scribe-iq-lakehouse — a production-pattern healthcare data lakehouse: a Bronze→Silver→Gold medallion over Synthea Coherent FHIR, built twice (Polars + delta-rs locally, Spark + Delta on Microsoft Fabric), orchestrated with Dagster, emitting one governed Gold data contract.
  • campus-rag-assistant — a multicloud RAG + agentic helpdesk platform: LangGraph orchestration, AWS Bedrock / Azure AI Search providers, cited answers, human-in-the-loop ticket filing, and RAGAS evals behind full CI/security gates.

Source & license