Documented result90% Recall@5100-query retrieval evaluation (local research; completed): Recall@5 for hybrid retrieval. 100 diagnostic queries; 20 service manuals; approximately 4,500 document chunks. Recall@5: 90%. Hybrid retrieval over 100 diagnostic queries; profile-supported, evaluation artifacts pending. Indexed document chunks: ~4,500. Approximate corpus size: 20 service manuals. Service manuals: 20. Indexed automotive service manuals. Decision-support POC, not a substitute for a certified technician. Recall describes the 100-query retrieval evaluation, not field diagnostic accuracy. Separate lexical/dense scores and scenario pass rates are not supplied.
Independent experiments across five EEG bands using 224×224 grayscale FRPs.
Documented result54.85% gamma · four-class segmentsFour-class segment-level Swin experiments (local research; completed): Four-class Swin baseline; segment-level per-band metrics. Specific split ratios and precision/recall/F1 averaging are not supplied. Selected TUSZ seizure recordings represented as 1-second FRP segments. Delta segment accuracy: 48.15%. Four classes (CPSZ/FNSZ/GNSZ/MYSZ); segment-level result, not patient generalization. Theta segment accuracy: 49.26%. Four classes (CPSZ/FNSZ/GNSZ/MYSZ); segment-level result, not patient generalization. Alpha segment accuracy: 46.17%. Four classes (CPSZ/FNSZ/GNSZ/MYSZ); segment-level result, not patient generalization. Beta segment accuracy: 48.78%. Four classes (CPSZ/FNSZ/GNSZ/MYSZ); segment-level result, not patient generalization. Gamma segment accuracy: 54.85%. Four classes (CPSZ/FNSZ/GNSZ/MYSZ); segment-level result, not patient generalization. Gamma segment precision: 47.32%. Four-class gamma-band segment experiment. Macro/weighted averaging is not supplied. Gamma segment recall: 47.49%. Four-class gamma-band segment experiment. Macro/weighted averaging is not supplied. Gamma segment F1: 46.96%. Four-class gamma-band segment experiment. Macro/weighted averaging is not supplied. Four-class segment discrimination remains challenging; accuracy is modest across bands. MYSZ has only one patient and is excluded from the patient-separated pilot. Only five held-out test patients; initial generalization pilot, not a definitive cross-patient performance estimate. Segment and patient results are not directly comparable: class sets, evaluation units and split strategies differ. Controlled CNN-vs-Swin patient-separated comparison is ongoing/planned; it has no supplied completed results. Internal Biomedical Sensors and Systems Lab, University of Memphis research; no public repository, publication URL or stable external artifact link is supplied.
Exploratory patient-separated pilot on only five held-out patients.
Documented result60% theta · three-class patient pilotThree-class patient-level generalization pilot (local research; completed): Patient-separated train/validation/test split; final patient prediction via majority voting. TUSZ patient pilot: 13 train, 3 validation and 5 held-out test patients. Delta patient accuracy: 40%. Three-class patient pilot (MYSZ excluded); five held-out patients only; not directly comparable to four-class segment results. Theta patient accuracy: 60%. Three-class patient pilot (MYSZ excluded); five held-out patients only; not directly comparable to four-class segment results. Alpha patient accuracy: 20%. Three-class patient pilot (MYSZ excluded); five held-out patients only; not directly comparable to four-class segment results. Beta patient accuracy: 20%. Three-class patient pilot (MYSZ excluded); five held-out patients only; not directly comparable to four-class segment results. Gamma patient accuracy: 20%. Three-class patient pilot (MYSZ excluded); five held-out patients only; not directly comparable to four-class segment results. Four-class segment discrimination remains challenging; accuracy is modest across bands. MYSZ has only one patient and is excluded from the patient-separated pilot. Only five held-out test patients; initial generalization pilot, not a definitive cross-patient performance estimate. Segment and patient results are not directly comparable: class sets, evaluation units and split strategies differ. Controlled CNN-vs-Swin patient-separated comparison is ongoing/planned; it has no supplied completed results. Internal Biomedical Sensors and Systems Lab, University of Memphis research; no public repository, publication URL or stable external artifact link is supplied.
7,641 chunks and 76 queries; recall, throughput, build time and reported memory overhead.
Documented result7,641 chunks · 76 queriesANN / exact retrieval benchmark (local research; completed): Unified VectorStore comparisons across ANN and exact/Flat approaches; RAGAS end-to-end quality evaluation is documented without supplied numerical scores. arXiv document corpus: 7,641 chunks, 76 queries. Document chunks: 7,641. Benchmark corpus size. Benchmark queries: 76. Number of retrieval benchmark queries. HNSW recall: ≥0.998. Local HNSW retrieval benchmark; recall cutoff and aggregation not supplied. HNSW retrieval throughput: ~1,439 QPS. Local retrieval queries per second, not application request latency or production throughput. FAISS-Flat index build time: 0.002 s. Local index construction time; not request latency. FAISS-Flat reported memory overhead: 0.09 MB. Reported local benchmark memory overhead; measurement definition not supplied. Local research application and retrieval benchmark, not a production service benchmark. QPS and index build time are not application request latency. Benchmark hardware, query distribution, recall cutoff and full per-backend outputs are not supplied.
VGG16 transfer learning with targeted fine-tuning; 94% validation accuracy across four categories.
Documented result94% validation accuracyFour-class model validation (local research; completed): Validation accuracy; split sizes and class labels are not supplied. Four poultry disease categories; dataset identity and sample count not supplied. Validation accuracy: 94%. Four-class validation accuracy; sample count and external validation not supplied. Validation accuracy across four classes; dataset size, split details, class names and external validation are not supplied. Sub-500 ms inference statement has no supplied hardware, percentile, sample size or production SLA definition.
EPCTL11 manual/YASA spindle calibration and cross-participant limits. Manual detector versus YASA detection counts on calibration participant EPCTL11. Cross-participant sensitivity and failure-mode inspection after EPCTL11 calibration.
Documented resultEPCTL11 manual/YASA spindle calibration and cross-participant limitsEPCTL11 spindle calibration (local research; completed): ANPHY-Sleep participants 11–15; Cz channel, 0.5–30 Hz broad-band and 11–16 Hz spindle-band filtering, Hilbert amplitude envelope, 100 ms smoothing, mean + 1.2 × standard deviation threshold, and 0.5–2.0 second duration constraints. EPCTL11 supplied the calibration comparison. ANPHY-Sleep calibration participant EPCTL11; count agreement, not annotated detection accuracy. Spindle cross-participant checks (local research; completed): Apply the calibrated manual detector and compare with cached YASA detections. EPCTL14 generalized well; EPCTL12 showed moderate over-sensitivity; EPCTL13 and EPCTL15 exposed failure modes. Manual detections generally had shorter mean durations. ANPHY-Sleep EPCTL12, EPCTL13, EPCTL14 and EPCTL15. Manual spindle detections: 484. EPCTL11 calibration only: 484 manual versus 468 YASA detections; difference +16, approximately +3.42% relative to YASA. More detections do not establish better accuracy. YASA spindle detections: 468. EPCTL11 calibration only: 484 manual versus 468 YASA detections; difference +16, approximately +3.42% relative to YASA. More detections do not establish better accuracy. Manual minus YASA detection count: +16. EPCTL11 calibration only: 484 manual versus 468 YASA detections; difference +16, approximately +3.42% relative to YASA. More detections do not establish better accuracy. Relative detection-count difference: ~+3.42%. EPCTL11 calibration only: 484 manual versus 468 YASA detections; difference +16, approximately +3.42% relative to YASA. More detections do not establish better accuracy. Research workflows and benchmarks, not clinical diagnostic systems or production medical devices. EPCTL11 detection-count agreement is a calibration-participant result, not universal detection improvement or validated accuracy, sensitivity or specificity. EPCTL14 generalized well; EPCTL12 showed moderate over-sensitivity; EPCTL13 and EPCTL15 exposed cross-subject failure modes. Manual detections generally had shorter average durations than YASA. No aggregate cross-participant quality metric is supplied.
109-subject audit establishes a clean 106-subject motor imagery baseline. 109 subjects audited; incompatible recordings excluded to establish a clean 106-subject baseline.
Documented result109-subject audit establishes a clean 106-subject motor imagery baselinePhysioNet sampling-frequency audit (local research; completed): Exclude S088, S092 and S100 (128 Hz rather than expected 160 Hz). Normalize channel names before montage assignment. PhysioNet EEG Motor Movement/Imagery cohort: 109 audited subjects, 3 exclusions, 106 retained for baseline. Audited subjects: 109. PhysioNet EEG Motor Movement/Imagery cohort: 109 audited subjects, 3 exclusions, 106 retained for baseline. Exclude S088, S092 and S100 (128 Hz rather than expected 160 Hz). Normalize channel names before montage assignment. Sampling-frequency exclusions: 3. PhysioNet EEG Motor Movement/Imagery cohort: 109 audited subjects, 3 exclusions, 106 retained for baseline. Exclude S088, S092 and S100 (128 Hz rather than expected 160 Hz). Normalize channel names before montage assignment. Clean baseline subjects: 106. PhysioNet EEG Motor Movement/Imagery cohort: 109 audited subjects, 3 exclusions, 106 retained for baseline. Exclude S088, S092 and S100 (128 Hz rather than expected 160 Hz). Normalize channel names before montage assignment. Research workflows and benchmarks, not clinical diagnostic systems or production medical devices. S088, S092 and S100 were excluded because their 128 Hz recordings differed from the expected 160 Hz; the clean baseline contains 106 of 109 audited subjects.
Baseline CSP classifiers achieve approximately 61.5% / 61.7% mean session-wise LOSO accuracy. Baseline CSP + LDA and CSP + linear SVM mean LOSO accuracy.
Documented resultBaseline CSP classifiers achieve approximately 61.5% / 61.7% mean session-wise LOSO accuracyMotor imagery baseline LOSO (local research; completed): 8–30 Hz filtering; cue-locked epochs -1.0 to 4.0 s, baseline -1.0 to 0.0 s, classification window 0.5 to 3.5 s. Fit CSP on training folds only. Subject-wise Leave-One-Session-Out across R04, R08 and R12, avoiding random trial splits. PhysioNet EEG Motor Movement/Imagery: clean 106-subject baseline, sessions R04, R08 and R12. Baseline CSP + LDA LOSO accuracy: ~61.5%. Baseline CSP + LDA; reported mean research LOSO accuracy. PhysioNet EEG Motor Movement/Imagery: clean 106-subject baseline, sessions R04, R08 and R12. Within-subject Leave-One-Session-Out (R04, R08, R12). Not a clinical or held-out-subject result. Baseline CSP + linear SVM LOSO accuracy: ~61.7%. Baseline CSP + linear SVM; reported mean research LOSO accuracy. PhysioNet EEG Motor Movement/Imagery: clean 106-subject baseline, sessions R04, R08 and R12. Within-subject Leave-One-Session-Out (R04, R08, R12). Not a clinical or held-out-subject result. Research workflows and benchmarks, not clinical diagnostic systems or production medical devices. S088, S092 and S100 were excluded because their 128 Hz recordings differed from the expected 160 Hz; the clean baseline contains 106 of 109 audited subjects. LOSO means Leave-One-Session-Out within subjects across R04, R08 and R12, not leave-one-subject-out or clinical performance. Tuning cohort size and uncertainty estimates are not separately supplied. Tuning yielded only about 1–1.2 percentage points of reported mean LOSO improvement; subject/session non-stationarity remained a larger constraint than classifier choice alone.
Tuned CSP classifiers achieve approximately 62.7% / 62.4% mean session-wise LOSO accuracy. Tuned CSP + LDA and CSP + linear SVM mean LOSO accuracy; no inferred tuning cohort size.
Documented resultTuned CSP classifiers achieve approximately 62.7% / 62.4% mean session-wise LOSO accuracyMotor imagery tuned LOSO (local research; completed): 8–30 Hz filtering; cue-locked epochs -1.0 to 4.0 s, baseline -1.0 to 0.0 s, classification window 0.5 to 3.5 s. Fit CSP on training folds only. Subject-wise Leave-One-Session-Out across R04, R08 and R12, avoiding random trial splits. LDA: 6 CSP components, LSQR, shrinkage = 0.3. Linear SVM: 8 CSP components. Confusion matrices, CSP spatial patterns and CSV summaries inspected in a Streamlit experiment dashboard. PhysioNet EEG Motor Movement/Imagery left-vs-right motor imagery; tuning cohort size not separately supplied. Tuned CSP + LDA LOSO accuracy: ~62.7%. 6 CSP components, LSQR and shrinkage = 0.3; reported mean research LOSO accuracy. PhysioNet EEG Motor Movement/Imagery left-vs-right motor imagery; tuning cohort size not separately supplied. Within-subject Leave-One-Session-Out (R04, R08, R12). Not a clinical or held-out-subject result. Tuned CSP + linear SVM LOSO accuracy: ~62.4%. 8 CSP components, linear SVM; reported mean research LOSO accuracy. PhysioNet EEG Motor Movement/Imagery left-vs-right motor imagery; tuning cohort size not separately supplied. Within-subject Leave-One-Session-Out (R04, R08, R12). Not a clinical or held-out-subject result. Research workflows and benchmarks, not clinical diagnostic systems or production medical devices. S088, S092 and S100 were excluded because their 128 Hz recordings differed from the expected 160 Hz; the clean baseline contains 106 of 109 audited subjects. LOSO means Leave-One-Session-Out within subjects across R04, R08 and R12, not leave-one-subject-out or clinical performance. Tuning cohort size and uncertainty estimates are not separately supplied. Tuning yielded only about 1–1.2 percentage points of reported mean LOSO improvement; subject/session non-stationarity remained a larger constraint than classifier choice alone.
Distributed PySpark ETL and aggregation on NYC Taxi data, benchmarked in reproducible single-node and multi-node Docker environments.
Documented result2.3× throughput gain in the compared multi-node setupSingle-node versus multi-node Spark benchmark (local research; completed): Reproducible Docker environments running PySpark cleaning, transformation and aggregation workflows; compared throughput, runtime and CPU/memory utilization alongside distributed-execution overhead. NYC Taxi dataset; dataset volume is not supplied. Multi-node relative throughput gain: 2.3×. Reported multi-node throughput relative to the compared single-node Spark setup on NYC Taxi data in a local Docker experiment. Absolute throughput, runtime, hardware, node count and dataset volume are not supplied; no production performance or cost savings are implied. Local distributed experiment, not a production data platform or cloud-managed Spark deployment. The reported 2.3× throughput gain applies only to the compared multi-node and single-node setups; it is not a general Spark performance guarantee. Dataset volume, cluster node count, worker configuration, hardware, absolute runtime values, cost figures and detailed benchmark tables are not supplied.