Abstract:Field service crews need routes respecting skills, shift limits, and customer time windows simultaneously. We combine QUBO-inspired tunneling moves with classical ruin-and-recreate neighborhoods and adapt penalty weights online when feasibility rates stall. Benchmark instances derived from municipal maintenance tickets achieve lower total tardiness than tabu search alone while staying within wall-clock budgets suitable for nightly replanning.
Abstract:Intermittent failures waste reviewer time when stack traces differ only in timing-dependent assertions. We cluster failing runs by templated log events and rank source lines whose coverage fluctuates with environment fingerprints such as container image digests. On six months of open-source CI exports, the ranking surfaces order-dependent database fixtures ahead of naive failure-frequency baselines.
Abstract:Narrowband UAV captures reveal chlorophyll stress earlier than RGB alone but labeled pixels are scarce each season. We pretrain masked autoencoders across unlabeled flights from multiple cooperatives and fine-tune shallow decoders with a handful of agronomist outlines. Cross-farm evaluation shows improved IoU on water-stressed canopies especially when sun glint patterns differ between pretraining and deployment weeks.
Abstract:Analytical views over append-only fact tables break when compaction rewrites files while dimension snapshots advance. We maintain versioned join sketches that reconcile late facts against frozen dimension boundaries and schedule partial re-aggregation only for affected partitions. Experiments on TPC-DS inspired workloads with injected arrival delays cut recomputation volume relative to full view refreshes without violating snapshot isolation guarantees.
Abstract:Extensions increasingly compile performance-critical paths to Wasm while retaining privileged APIs. We derive fine-grained CFI labels from structured control-flow graphs and enforce them with a thin runtime monitor injected at link time. On a corpus of open-source shopping and password-manager extensions, the policy blocks synthetic return-oriented gadget chains with sub-five-percent runtime overhead on microbenchmark suites.
Abstract:Authors publish statistical supplements as wide spreadsheets that export to inaccessible HTML grids. We infer header scopes, units, and summary roles using constraint propagation over cell formatting cues and publish ARIA-enriched tables. User studies with blind experts report faster navigation on government open-data releases compared to default export plugins when merged cells and multi-row headers are present.
Abstract:Enterprise deployments rotate validators for compliance yet still require safety during membership waves. We extend a practical BFT pipeline with staggered view changes and buffered client certificates that survive brief quorum gaps. Trace-driven simulations using recorded churn schedules from a regional trade-finance pilot show fewer forced client retries than static quorum configurations at comparable median commit latency.
Abstract:Serving specialized assistants on modest GPU pools benefits from accepting draft tokens only when their short-horizon distributions align with the target model. We fit lightweight histogram calibrators on in-domain corpora and gate draft acceptance with entropy-aware thresholds. On legal and biomedical held-out prompts, calibrated speculation raises tokens per second without increasing exact-match disagreement rates relative to uncalibrated baselines.
Abstract:Colocation incentives push operators toward denser racks but migration churn wastes joules when demand shifts hourly. We formulate placement as an online problem with penalties for pod rescheduling and apply a regularized follow-the-regularized-leader update on latent CPU-memory traces. Simulated cluster traces show reduced cumulative energy plus switching cost against greedy bin-packing heuristics during diurnal load swings.
Abstract:Campaign-style accounts exploit layered follow and retweet relationships that single-layer metrics miss. We construct a multiplex graph with lagged edges and mine recurring three-node temporal motifs whose burst rates diverge from organic baselines estimated via block bootstrap. The ranking improves precision at fixed review budget relative to volume-only anomaly scores on labeled incident snapshots.