Adithya Parthasarathy

Regulated personalization systems must decide what information, offer, limit, explanation, or next action to present while preserving privacy and controlling downstream harm. Conventional sequential recommenders optimize engagement over behavior traces, but regulated settings also require privacy budgets, evidence constraints, risk limits, and auditable tool use. This paper proposes Privacy-Preserving Sequential Decision Systems (PPSDS), a synthetic architecture for regulated personalization that combines efficient sequence modeling, differentially private behavior summaries, constrained distributional policy gradients, hybrid evidence retrieval, distributed RAG, and contract-bound agent interoperability. PPSDS extends Ideological Drift Detection in Governed Enterprise Knowledge Bases with a drift sentinel for personalization narratives and extends Risk-Aware Financial RAG with Distributional Retrieval Policies with action-level privacy and tail-risk gates. In a simulated personalization benchmark spanning finance, benefits, and customer-support recommendations, PPSDS preserves most of the utility of an unconstrained sequential policy while reducing simulated privacy exposure by 72.0%, lowering conditional value-at-risk of adverse outcomes by 41.5%, and eliminating contract-violating evidence accesses.

Wide-area grid monitoring increasingly depends on high-rate phasor measurement unit (PMU) streams, cloud-native telemetry fabrics, predictive calibration models, and operator-facing decision support. These capabilities are usually deployed as separate systems: PMU analytics detect events, maintenance models estimate instrument risk, forecasting models project future load or stability margins, and schedulers route workloads across edge and cloud resources. This paper proposes an Edge-to-Cloud PMU Digital Twin (ECPDT), a synthetic architecture that binds those functions into one governed control loop. ECPDT maintains a continuously updated grid-state twin from PMU streams, predicts calibration time-to-degradation, forecasts long-horizon operating envelopes, and recommends bandwidth, placement, inspection, and control-support actions under policy and latency constraints. The design extends Low-Latency Grid Intelligence with Self-Governing Stream and Calibration Agents by adding a persistent digital-twin state and extends Cross-Cloud LLMOps Scheduler for Privacy-Budgeted RAG and Inference by using provider-aware scheduling for grid analytics and model inference. A simulated evaluation over synthetic PMU disturbances, calibration drift, and cross-cloud workload variation suggests that ECPDT improves event localization F1 from 0.72 to 0.89, reduces calibration-warning lead-time error by 37.5%, and lowers P95 decision-support latency by 34.1% relative to separated edge and cloud baselines.

Enterprise knowledge bases encode more than facts. They also encode policy preferences, operational assumptions, persuasive phrases, reviewer norms, and routing defaults that shape downstream retrieval-augmented generation and recommendation behavior. This paper proposes Ideological Drift Detection for Governed Knowledge Bases (IDD-GKB), a synthetic architecture for detecting when recurring policy narratives, compliance phrases, documentation claims, and retrieved evidence gradually change meaning or influence while remaining superficially consistent. IDD-GKB combines phrase-lineage analysis, hybrid semantic-relational retrieval, critic-guided documentation maintenance, distributed RAG, anonymized evidence views, and latency-aware sequence models. It extends Retrieval-Grounded Documentation Agents for Enterprise Compliance Evidence by adding drift scores over claim lineages and extends Cross-Cloud LLMOps Scheduler for Privacy-Budgeted RAG and Inference by feeding those drift scores into privacy-budgeted route and context selection. We define the architecture, a lineage-and-evidence drift model, and a simulated benchmark over governed policy, support, incident, and compliance corpora. In simulation, IDD-GKB improves drift-event F1 from 0.61 to 0.84, reduces unsupported narrative propagation by 57.8%, and preserves zero unauthorized sensitive-field disclosures under the benchmark privacy policy.

Enterprise compliance documentation is often written after systems have already changed, which leaves policy narratives, architecture diagrams, runbooks, and audit evidence out of sync with executable behavior. This paper proposes Retrieval-Grounded Documentation Agents (RGDA), a synthetic architecture in which bounded documentation agents maintain compliance text by combining abstract-syntax-tree-aware reflexion, hybrid semantic-relational evidence packs, distributed retrieval-augmented generation, model-context-protocol schemas, anonymized evidence views, and legal-meme lineage scoring. RGDA extends the policy-verified governance fabric of Policy-Verified Agentic DataOps for Regulated Multi-Cloud Analytics and the telemetry-to-evidence loop of Low-Latency Grid Intelligence with Self-Governing Stream and Calibration Agents by treating documentation edits as governed actions that must cite executable, operational, and policy evidence before they can be accepted. We define the architecture, evidence model, documentation-maintenance algorithm, and a simulated compliance benchmark over software, cloud, and grid-observability workflows. In simulation, RGDA improves evidence-supported claim accuracy from 73.2% to 94.8%, reduces stale compliance claims by 64.7%, and eliminates unauthorized sensitive-evidence exposure relative to an unbounded documentation-agent baseline. These results are hypothetical and are intended to make the design auditable rather than to claim production deployment.

Wide-area power-grid observability depends on phasor measurement unit (PMU) streams, calibration-aware instrumentation, and fast operator workflows, but the analytic stack that joins these elements is often split between stream processors, model-serving systems, privacy filters, and manual maintenance queues. This paper proposes Self-Governing Grid Intelligence (SGI), a synthetic cyber-physical stream-processing architecture in which bounded agents coordinate PMU telemetry, calibration-risk transformers, long-horizon forecasts, anonymization controls, and model-context-protocol contracts under policy-verified cloud-pipeline governance. SGI extends the policy-verified agentic DataOps fabric of S01 to the grid domain by adding sub-second latency budgets, measurement-quality gates, calibration risk scoring, and maintenance-action escalation. We define the architecture, a latency-and-risk control algorithm, and a simulated benchmark over wide-area disturbance monitoring and instrument maintenance scenarios. In the synthetic benchmark, SGI reduces P95 event-to-decision latency by 61.5% relative to a batch-calibrated PMU baseline while improving calibration-risk F1 from 0.68 to 0.83 and eliminating unauthorized privacy-policy violations. These results are simulated and should be interpreted as an executable design target rather than evidence from a production utility deployment.

This paper studies a simplified instance of the design question raised by GPU A-SGD: how should stochastic optimization be organized so that device-level parallelism improves throughput without destabilizing convergence? Motivated by the asynchronous multi-GPU training framework of Paine et al., we examine a controlled linear-regression setting in which workers process disjoint data shards, compute stochastic gradient updates in parallel, and combine local models through either plain averaging or a lightweight asynchronous consensus rule. The resulting study isolates three effects that matter in larger systems as well: the runtime crossover between CPU and GPU execution, the point at which GPU utilization saturates despite transfer overhead, and the robustness benefit of limited cross-worker coordination when data variability increases. The measurements show a clear scale threshold beyond which GPU execution becomes favorable, near-saturated utilization on modestly sized workloads, and consistently lower root-mean-square error for the asynchronous variant in higher-variance regimes. These findings support the broader claim of the GPU A-SGD literature: parallel stochastic optimization is most effective when compute throughput and communication discipline are designed together.

This paper studies automatic attribute extraction and sentiment assignment from product reviews. The task is to recover product aspects from free-form customer reviews and label them as positive, negative, or irrelevant. The modeling progression runs from two lightweight heuristics to two discriminative sequence models: a noun-driven bag-of-words baseline, a nearest-adjective sentiment heuristic, a Maximum Entropy Markov Model (MEMM), and a Conditional Random Field (CRF). The empirical pattern is clear. Heuristic models are adequate for high-recall noun harvesting and for recognizing the majority NONE class, but they struggle to assign polarity reliably because sentiment is often non-local and multiple aspects can appear in the same sentence. Sequence models improve substantially by conditioning on lexical, orthographic, part-of-speech, chunk, and contextual cues. Feature-level F1 rises into the high 90s for MEMM and CRF, while polarity-specific F1 remains meaningfully harder, especially under cross-domain transfer. The strongest overall conclusions are that language structure matters more than sentence-level sentiment alone, and that future gains will come from better modeling of multi-word aspects, longer-range dependencies, and domain shift.

Regulated analytics teams increasingly run data pipelines, retrieval systems, and machine learning workloads across multiple cloud providers, but the operational control plane remains split across pipeline schedulers, API gateways, privacy tooling, and audit systems. This paper proposes Policy-Verified Agentic DataOps (PVA-DataOps), a synthetic reference architecture in which bounded agents propose operational actions while a policy verifier, evidence layer, and contract-mediated gateway decide whether those actions may execute. The design combines governed cloud pipeline agents, intent-aware multi-cluster gateways, cross-cloud workload routing, hybrid semantic-relational retrieval, distributed retrieval-augmented generation (RAG), model-context contracts, and privacy-preserving anonymization. We describe the control architecture, a verification algorithm, and a simulated benchmark covering recovery time, policy drift, evidence validity, privacy-utility loss, and cost. In simulation, PVA-DataOps reduces mean recovery time by 46.3% relative to static orchestration while eliminating unapproved policy-drift events. The results are synthetic and are intended to make the architecture testable rather than to claim production deployment.

Forecasting when reviewers will agree or disagree is useful for peer-review management because high-disagreement submissions often require additional discussion, stronger area-chair intervention, or more careful reviewer assignment. Prior computational work has mostly focused on predicting final decisions or review scores from papers and completed reviews, while a separate legal machine learning line introduced citation-propagated phrase scoring to model agreement through "memes" that spread over a citation network. This paper proposes a modest empirical-lite adaptation of that idea to scientific peer review. Each reviewer is represented by a historical profile built only from information available before the target submission, with two complementary views: a citation-grounded phrase genealogy inspired by Verma et al. and a dense semantic profile derived from scientific document encoders. Pairwise reviewer agreement is then predicted from reviewer-reviewer similarity, reviewer-submission affinity, and a small set of chronology-safe metadata features. A public-data proof of concept on open F1000Research reviews from RottenReviews shows that the pipeline is feasible and yields a modest macro-F1 lift over a majority baseline, but ranking discrimination remains near chance and calibration remains weak. The main takeaway is therefore intentionally limited: historical phrase profiles are implementable and worth studying further, but the current public-data proxy does not yet establish strong predictive evidence.

Predicting disagreement on appellate panels is a useful problem for empirical legal studies because dissent is rare, institutionally important, and often tied to persistent differences in judicial style and precedent use. Prior English-language work has shown that structured case features, seating patterns, and citation-derived phrase statistics can help predict vote alignment in the U.S. Courts of Appeals, while recent legal NLP work has introduced domain-specific language models and stronger long-document encoders. This paper proposes a modest empirical synthesis of those lines of work. The key idea is to retain the meme-scoring intuition of Verma et al., but compute meme and semantic profiles only from each judge's historical opinions prior to the target case. Pairwise agreement is then predicted from structured panel-history features, historical meme overlap, and legal language-model similarity. A public-data pilot on streamed Caselaw Access Project opinions evaluates dissent-derived disagreement labels under a chronological split. On 1,764 held-out judge-pair examples with 173 disagreement positives, the hybrid model reaches macro F1 of 0.588, disagreement-class F1 of 0.335, average precision of 0.255, and ROC AUC of 0.772. These results are not definitive vote-label evidence, but they suggest that historical judicial profiles contain measurable ranking signal for future panel disagreement.

Semi-supervised image classification seeks to improve predictive accuracy by combining a small labeled set with a larger unlabeled pool. Currently, deep semi-supervised learning has produced strong benchmark results through entropy-based objectives, pseudo-labeling, self-ensembling, and perturbation consistency, yet a practical tension remains: methods that use every unlabeled example can absorb harmful targets, while simple pseudo-labeling can be unstable when model confidence is miscalibrated. This paper proposes a modest empirical-lite method, Agreement-Gated Self-Ensembling (AGSE), that combines an exponential-moving-average teacher with a conservative pseudo-label gate based on calibrated confidence, cross-view class agreement, and low prediction divergence. The proposal is intended for closed-set image classification on small-image benchmarks such as CIFAR-10 and SVHN, and is framed as a lightweight extension of existing self-ensembling and pseudo-labeling techniques. A synthetic pilot comparison is included only as placeholder evidence for paper layout and narrative planning; it suggests that selective pseudo-label acceptance could improve label efficiency without requiring a generator, strong augmentation policy, or large architectural change. The main takeaway is that a carefully restricted bridge between self-ensembling and pseudo-label selection is a plausible and appropriately modest research contribution for the current semi-supervised learning literature.

This paper presents a two-stage forecasting framework for future carbon dioxide (CO2) emissions using historical country-level indicators. The method first extrapolates the future trajectories of the input indicators and then predicts future CO2 emissions with nonlinear regressors trained on the most recent segment of the time series. Two design choices define the approach: recency-weighted training, which increases the influence of modern observations, and a boundary-consistency rule that selects the forecast with the smoother transition from observed to predicted emissions. The resulting forecasts preserve country-level heterogeneity while producing a globally increasing aggregate CO2 trajectory, with reported agreement against an external scientific baseline of 95.4%. The accompanying empirical results also show that the strongest associated indicators are dominated by energy use, fossil-fuel electricity production, transport activity, and trade-related variables.