DevOps & Infra

ndpvt-web Skill

Build graph-grounded multihop QA systems over regulatory and hierarchically structured documents. Combines vector similarity retrieval with knowledge graph traversal to answer questions requiring reasoning across multiple interconnected legal or policy documents. Triggers: 'build a regulatory QA system', 'multihop question answering over legal documents', 'graph-based retrieval for regulations', 'answer questions across multiple policy documents', 'healthcare regulation QA pipeline', 'cross-document reasoning with graph expansion'

ndpvt-web Skill

Build intelligent alert lifecycle management systems for cloud infrastructure using graph-based denoising, RAG-powered summarization, and multi-agent rule refinement. Trigger phrases: - "reduce alert fatigue in our monitoring system" - "deduplicate and correlate alerts" - "summarize alerts for on-call engineers" - "refine our alerting rules automatically" - "build an alert denoising pipeline" - "too many alerts, help me triage"

ndpvt-web Skill

Apply CRAFT (Calibrated Reasoning with Answer-Faithful Traces) for multi-hop question answering with verified reasoning chains. Use when: 'answer a complex question from multiple documents', 'build a faithful multi-hop QA pipeline', 'reason over retrieved passages with citations', 'verify reasoning chain faithfulness', 'structured RAG with auditable traces', 'multi-step question decomposition with evidence grounding'

ndpvt-web Skill

Design production-grade agentic AI architectures with separated cognition/execution layers, typed tool interfaces, multi-agent topologies, and enterprise hardening. Use when: 'design an agent system', 'build a multi-agent architecture', 'add governance to my AI pipeline', 'harden my LLM agent for production', 'create a tool registry for agents', 'architect agent-to-agent coordination'.

ndpvt-web Skill

Build multi-agent pipelines that use knowledge graphs to guide LLM reasoning across domains. Agents specialize in problem decomposition, evidence retrieval, parameter extraction, graph traversal, and hypothesis synthesis. Use when: 'build a knowledge graph agent pipeline', 'find cross-domain material substitutes', 'multi-agent graph traversal system', 'knowledge graph guided search', 'graph-based RAG with specialized agents', 'cross-domain design exploration with KG'.

ndpvt-web Skill

Fact-check claims using knowledge graph-based contrastive reasoning. Constructs a KG from claims and evidence sources, generates contrastive questions ("Why X rather than Y?") grounded in the KG structure, distills evidence through targeted Q&A, and synthesizes a summary for veracity assessment. Trigger phrases: "fact-check this claim", "verify this statement against sources", "build a knowledge graph for claim verification", "contrastive reasoning for fact-checking", "check if this claim is true using these reports", "KG-CRAFT pipeline"

ndpvt-web Skill

Structure database schemas and structured data as file-native context for LLM agent operations. Applies evidence-based format selection, domain-partitioned file layouts, and architecture-aware context engineering to maximize accuracy when agents generate SQL, query APIs, or navigate large structured systems. Use this skill when: - "Generate SQL against a database with many tables" - "Set up schema context for an agent pipeline" - "Organize database metadata files for LLM consumption" - "Help me structure context for a text-to-SQL system" - "Scale my agent to handle thousands of tables" - "Choose the best format for feeding schema info to an LLM"

ndpvt-web Skill

Build multi-agent pipelines that generate realistic synthetic patient profiles by integrating epidemiological data, medical claims, literature evidence, and personality models. Use when asked to 'simulate patients', 'generate synthetic medical data', 'build a patient simulation pipeline', 'create virtual patient cohorts', 'model disease progression with agents', or 'synthesize clinical profiles'.

ndpvt-web Skill

Design scalable tool retrieval systems using hierarchical code tokenization that captures collaborative tool semantics. Use when: 'build a tool registry with hierarchical codes', 'scale tool selection for thousands of APIs', 'encode tool co-usage patterns', 'design a generative tool retrieval pipeline', 'organize APIs into a searchable hierarchy', 'implement constrained decoding for tool selection'.

ndpvt-web Skill

Build full-duplex voice dialogue systems using unit-based agent decomposition and semi-cascaded pipelines. Trigger phrases: 'build a full-duplex dialogue system', 'implement voice interaction with interruption handling', 'create a real-time conversational agent', 'design a semi-cascaded speech pipeline', 'add backchannel and turn-taking to a voice agent', 'unit-based dialogue processing'

ndpvt-web Skill

Build ROI-guided perturbation pipelines for explainable medical image segmentation using CLIP embeddings. Generates boundary-aware saliency maps by restricting perturbations to anatomically meaningful regions identified via vision-language models. Use when: 'explain medical segmentation predictions', 'build XAI pipeline for CT/MRI', 'generate saliency maps for organ segmentation', 'reduce perturbation cost for explainability', 'CLIP-guided occlusion sensitivity', 'ROI-aware LIME/RISE for medical images'.

ndpvt-web Skill

Structure-aware document reasoning that converts PDFs/long documents into hierarchically indexed paragraphs with coordinate metadata, then uses a dual-tool "locate then read" strategy (Retrieve + ReadSection) to answer complex questions requiring evidence scattered across distant document sections. Trigger phrases: - "analyze this PDF and answer questions about it" - "find evidence across sections of this document" - "search this long document for specific information" - "extract structured answers from this report" - "answer questions about this paper/filing/manual" - "build a document QA pipeline"

ndpvt-web Skill

Build grounded OCR pipelines using GutenOCR's prompt-based interface for reading, detection, and spatial grounding on documents. Use when: 'extract text with bounding boxes from a PDF', 'find where a phrase appears in a scanned document', 'build a document OCR pipeline with spatial coordinates', 'detect text regions in business documents', 'locate specific content in a scientific paper image', 'read text from a cropped region of a document'.

ndpvt-web Skill

Memory-enhanced multi-agent retrieval and reasoning for veracity assessment and fact-checking. Use when: 'verify this claim', 'fact-check these statements', 'check if this is true', 'assess the veracity of', 'cross-reference these claims', 'build a fact-checking pipeline'.

ndpvt-web Skill

Implement boundary-aware evidence selection for RAG systems using the BAR-RAG technique. Replaces relevance-only reranking with difficulty-calibrated evidence selection that targets the generator's competence boundary -- passages that are challenging yet solvable. Use when: "build a robust RAG pipeline", "improve reranking for noisy retrieval", "select better evidence for question answering", "make my RAG system handle noisy documents", "implement boundary-aware reranking", "calibrate evidence difficulty for my LLM generator".

ndpvt-web Skill

Apply the SAGE self-hinting technique to improve LLM problem-solving by generating graduated hints that boost solution diversity and prevent reasoning collapse. Use when: 'help me solve this hard math problem step by step', 'generate hints for this coding challenge', 'break down this complex problem with progressive clues', 'design a self-hinting prompt pipeline', 'create a hint curriculum for training data', 'scaffold this reasoning task with decomposition hints'.

ndpvt-web Skill

Build LLM-based pipelines that automatically detect and classify longitudinal (temporal) changes in radiology reports. Use when the user mentions 'radiology report annotation', 'longitudinal report evaluation', 'temporal change detection in medical text', 'disease progression extraction', 'radiology NLP pipeline', or 'benchmark radiology report generation'.

ndpvt-web Skill

Implement VTC-R1 vision-text compression for efficient long-context reasoning. Renders intermediate reasoning segments into images as 'optical memory' fed back into vision-language models, achieving 3.4x token compression and 2.7x latency speedup. Use when: 'compress long reasoning chains', 'implement optical memory for VLMs', 'build iterative vision-text reasoning loop', 'reduce inference tokens with image rendering', 'VTC-R1 reasoning pipeline', 'text-to-image compression for LLM reasoning'.

ndpvt-web Skill

Audit LLM agent systems for trust-boundary security vulnerabilities using the AgentFence taxonomy of 14 attack classes across planning, memory, retrieval, tool use, and delegation. Produces trace-auditable security reports with mean security break rates (MSBR) per attack surface. Use when: "audit my agent for security vulnerabilities", "check agent trust boundaries", "find delegation attacks in my agent code", "map security risks in my LangGraph/CrewAI pipeline", "evaluate agent architecture security", "run AgentFence analysis on this agent system".

ndpvt-web Skill

Structured multi-agent courtroom debate for explainable high-stakes tabular decisions. Use when: 'set up a multi-agent debate for this prediction', 'use adversarial agents to classify this table row', 'build a courtroom-style reasoning pipeline', 'create prosecutor/defense/judge agents for this decision', 'explain this tabular prediction with a debate', 'run a structured adversarial analysis on this dataset'.

ndpvt-web Skill

Build sequential LLM pipelines for scientometric question answering over academic databases. Decomposes meta-scientific queries into entity recognition, multi-step planning, parallel data retrieval, and analytical synthesis. Use when: 'build a scientometric QA system', 'answer questions about research impact', 'query academic publication databases with natural language', 'analyze citation metrics for institutions or authors', 'create a pipeline for science-of-science questions', 'implement RAG over scholarly metadata'.

ndpvt-web Skill

Build robust multi-tool agentic pipelines with noise-aware execution, parallel reasoning, and environment scaling patterns from the LongCat-Flash-Thinking architecture. Use when: 'build a multi-tool agent pipeline', 'make my agent robust to API failures', 'scale tool orchestration across domains', 'add parallel reasoning to my agent', 'handle noisy tool responses', 'design a fault-tolerant agentic workflow'.

ndpvt-web Skill

Design and implement process verification for multi-agent LLM systems. Add intermediate-step evaluation to multi-agent workflows using LLM-as-a-Judge, reward models, or process reward models at agent-level or iteration-level granularity. Trigger phrases: 'add verification to my multi-agent system', 'evaluate intermediate steps in my agent pipeline', 'process verification for MAS', 'verify agent reasoning steps', 'judge agent outputs in a multi-agent workflow', 'add quality checks between agent steps'.

ndpvt-web Skill

Orchestrate multi-agent LLM pipelines using constraint factorization -- decomposing complex requirements into separate constraint-enforcement agents that iteratively project a shared solution toward feasibility. Use when: 'build a multi-agent pipeline', 'decompose this problem into agent constraints', 'set up constraint-based agent orchestration', 'factorize requirements across agents', 'iterative multi-agent refinement loop', 'cyclic projection agent workflow'.

ndpvt-web Skill

Build collaborative multi-agent systems that use alternating reasoning-reflection cycles with specialized agent roles, coarse-to-fine selection, dynamic focus layouts, and chain-of-reflection verification. Use when asked to: 'build a multi-agent pipeline with self-verification', 'create agents that reflect on their own outputs', 'design a coarse-to-fine selection system', 'implement a questioner-responder verification loop', 'build a reasoning pipeline with iterative refinement', 'create a multi-step agent workflow with feedback loops'.

ndpvt-web Skill

Implement predictive safety guardrails for computer-using agents and automated pipelines using world-model-based risk prediction. Catches long-term risks that reactive checks miss by simulating future states before executing actions. Use when: 'add safety guardrails to my agent', 'predict risks before executing shell commands', 'build a predictive safety layer', 'guardrail my automation pipeline', 'prevent irreversible actions in my agent workflow', 'add a world model safety check to my CUA'.

ndpvt-web Skill

Systematic root cause analysis for cloud/distributed system failures using a 16-category reasoning failure taxonomy and multi-hop fault propagation tracing. Use when: 'diagnose this production incident', 'find the root cause of this outage', 'trace this failure through our microservices', 'why is this service failing', 'analyze these alerts and logs to find what broke', 'RCA for this distributed system issue'.

ndpvt-web Skill

Build scope-matched knowledge graph RAG pipelines where retrieval precision beats breadth. Constructs domain-specific KGs from scientific literature, selects scope-aligned subgraphs for retrieval, and injects focused context into LLM prompts — avoiding the accuracy loss caused by union-based retrieval. Use when: "Build a knowledge graph RAG pipeline for medical questions", "Add domain-specific retrieval to my LLM app", "My RAG pipeline returns too much irrelevant context", "Help me scope my knowledge graph to match my query domain", "Design a biomedical QA system with knowledge graphs", "Reduce noise in my retrieval-augmented generation system"

ndpvt-web Skill

Design and implement compiler-driven hierarchical memory offloading for LLM inference and training on multi-tier memory systems. Applies graph-level scheduling of data movement to hide memory transfer latency behind computation. Use when: 'optimize LLM memory offloading', 'schedule prefetch and evict in computation graph', 'hierarchical memory management for inference', 'hide data transfer latency in ML pipeline', 'compiler pass for memory placement', 'reduce peak GPU memory with offloading'.

ndpvt-web Skill

Implement JitRL-style continual learning for LLM agents: training-free policy optimization via experience memory, advantage estimation, and logit modulation. Use when asked to 'add experience memory to an agent', 'implement continual learning without fine-tuning', 'build a JitRL agent', 'optimize agent actions from past trajectories', 'add non-parametric RL to an LLM pipeline', or 'make my agent learn from its mistakes at inference time'.

ndpvt-web Skill

Debug multi-step tool-using agent pipelines by localizing the first irrecoverable error via binary-search rollback, then concentrating fixes on that critical step and its downstream suffix. Use when: 'find where my agent pipeline first breaks', 'debug this tool chain failure', 'locate the root cause in my multi-step workflow', 'which step killed this reasoning chain', 'binary search for the breaking step', 'error-localize my agent trajectory'.

ndpvt-web Skill

Build reviewer/expert recommendation systems using LLM-generated semantic profiles and Multi-gate Mixture-of-Experts (MMoE) multi-task learning. Applies the OmniReview Pro-MMoE architecture for matching documents to qualified reviewers or experts. Use when: 'recommend reviewers for this paper', 'build an expert matching system', 'find qualified reviewers from a candidate pool', 'rank experts by relevance to a submission', 'build a reviewer recommendation pipeline', 'match papers to domain experts'.

ndpvt-web Skill

Build retrieval-augmented in-context learning (RAICL) pipelines that convert time-series or signal data into images and classify them with vision-language models using dynamically retrieved few-shot examples. Trigger phrases: 'classify EEG with a VLM', 'retrieval-augmented in-context learning', 'convert signals to images for GPT/Gemini', 'few-shot VLM classification', 'RAICL pipeline', 'time-series image classification with LLM'

ndpvt-web Skill

Audit and harden AI/ML systems for reliability and responsibility using an 8-dimension framework covering bias, security, privacy, hallucination, uncertainty, explainability, distribution shift, alignment, and AIGC detection. Use this skill when users say things like: - "Audit my LLM application for safety and fairness" - "Add responsible AI checks to my ML pipeline" - "How do I detect hallucinations in my model output?" - "Harden my AI system against adversarial attacks" - "Add bias detection to my NLP pipeline" - "Implement watermarking for AI-generated content"

ndpvt-web Skill

Evaluate robustness of LLM-generated recommendation explanations under realistic user behavior noise. Use when: 'test explanation robustness', 'evaluate recommender explanation stability', 'perturb user history and measure explanation drift', 'benchmark LLM explanation consistency', 'build a robustness evaluation pipeline for recommendation agents', 'stress-test explanation generation under noisy inputs'.

ndpvt-web Skill

Implement uncertainty quantification and calibration assessment for LLM-generated long-form answers. Apply answer-frequency consistency, verbalized confidence elicitation, token-level analysis, and multi-metric calibration benchmarking based on the UQ framework from Müller et al. (2026). Trigger phrases: - "measure how confident the model is in this answer" - "calibrate uncertainty on these QA results" - "benchmark uncertainty quantification for my LLM pipeline" - "which uncertainty method should I use for scientific QA" - "detect unreliable LLM answers" - "evaluate calibration of model confidence scores"