An open-source edtech infrastructure project built production-first. PLRS combines a SAKT (Self-Attentive Knowledge Tracing) model trained on real student interaction data, a knowledge graph constraint layer that enforces prerequisite ordering, and a multi-objective ranker that incorporates SuperMemo-2 spaced repetition as a fourth signal. The result is a recommendation system where prerequisite violations drop to zero — compared to over 81% for conventional collaborative filtering baselines.
Semantic job matching pipeline using LLM embeddings and vector search. Retrieves and ranks listings against candidate profiles with explainable scoring and source attribution.
↗End-to-end time-series system for commodity and market data. Automated feature engineering, ensemble models, and live prediction endpoints with drift monitoring.
↗Production data pipeline with schema validation, incremental loading, and alerting. Multi-source ingestion with lineage tracking and idempotent transforms.
↗Orchestrated agent system for automated research and summarization. Tool-use, memory, and structured output with evaluation harness for response quality.
Diagnosed data leakage in provided aggregate columns — identified as the primary driver of leaderboard score gaps — then produced a diagnostic pipeline and calibrated submission leveraging this understanding. A good reminder that understanding your data deeply matters more than stacking models.