Beyond the Model: Workshop on Machine Learning Systems, Federated Learning, and MLOps

Cross-silo federated learning still needs mechanisms for participant privacy and accountability. PBS-FL separates client authentication from update submission using partially blind credentials and round-specific anonymous identifiers on a permissioned blockchain, accepting at most one contribution per authorized client per round. A lightweight Tally Hash lets a client confirm its update was part of the aggregation input. The prototype runs on Hyperledger Fabric and was evaluated on two credit-risk datasets: predictive accuracy is unchanged, anonymous authorization adds bounded overhead, and Tally Hash generation stayed under one millisecond.

The authors re-evaluate their earlier kNN + SVD stacking ensemble for recommending programming problems against implicit ALS, Neural Collaborative Filtering, and LightGCN, all under one automated pipeline: temporal splits, fixed-seed evaluation, parallel CI training, an MLflow registry with champion selection, and drift monitoring that triggers retraining. The two datasets give opposite verdicts. On CodeChef a popularity baseline goes unbeaten and only LightGCN matches it statistically; on Aizu popularity finishes last and implicit ALS wins decisively. A catalog-subsampling control traces the contrast to the temporal stability of item popularity (PSI 0.02 versus 0.72) rather than catalog size. The original hybrid stack is uncompetitive on both.

Most RAG pipelines use a single similarity metric both to retrieve context and to certify that the context is relevant, which ties retrieval speed to retrieval trust. IAssist-QCI, a RAG-based institutional assistant, splits the two: a retrieval path optimized for fast candidate fetching over chunk-level vector similarity, and a verification path that independently computes a composite relevance score over the whole multi-chunk context injected into the prompt. The split is what makes two further capabilities possible: low verification scores become automated knowledge-gap alerts that detect drift in the knowledge base without manual auditing, and an asynchronous curation pipeline builds a synthetic question-answer dataset for later fine-tuning without interrupting production traffic. Evaluation uses relevance-score distributions and gap-alert rates from 30 days of operation.

Three years of operating a Performance Ratio monitoring pipeline for seven first-generation 1 MW AC photovoltaic plants in southern Ecuador, stitched across five incompatible manufacturer platforms under limited infrastructure and regulatory constraints on data sharing. Two findings drive the report. A statistically functional conditional-percentile model was retired in favour of a deterministic physical estimator because only the latter is defensible before the national regulator, so auditability decided the method, not accuracy. Then a ten-stage audit of the running system found that the self-reference flaw identified in the retired model had reappeared unnoticed in the alerting rule that replaced it: injecting degradation into real series, the moving baseline absorbs 7.7 percentage points of decay and never fires below 3%/month. The report also documents a capacity denominator with no recorded provenance on all seven plants, and draws six transferable lessons.