DevOps & Infra

ndpvt-web Skill

Generate and evaluate visual metonymy -- indirect visual representations that evoke concepts through associated cues rather than literal depiction. Uses a semiotic-theory-grounded pipeline (representamen generation, chain-of-thought visual description, image synthesis) to create images where meaning is implied, not shown. Trigger phrases: 'generate visual metonymy', 'create indirect visual representation', 'visual metaphor pipeline', 'metonymic image generation', 'semiotic image prompt', 'evoke concept visually without showing it'.

ndpvt-web Skill

Build RAG pipelines that transform unstructured clinical or domain-specific documents into structured form records using a multi-layer data lake, hybrid relational+vector storage, and rule-driven adaptive forms. Trigger phrases: 'build a clinical document extraction pipeline', 'convert unstructured reports to structured forms', 'RAG pipeline for medical records', 'automate form completion from documents', 'extract structured data from clinical notes', 'build a tumor board automation system'.

ndpvt-web Skill

Build contamination-aware hybrid RAG evaluation pipelines that couple knowledge graphs with text retrieval for multi-hop reasoning benchmarks. Use when: 'build a RAG benchmark', 'evaluate multi-hop reasoning', 'create hybrid KG-text retrieval pipeline', 'detect parametric recall vs genuine reasoning', 'generate multi-hop QA from knowledge graphs', 'benchmark RAG system with contamination control'.

ndpvt-web Skill

Self-evolving multi-agent orchestration that dynamically generates specialized roles and collaboration topologies at inference time. Instead of fixed agent roles, MetaGen creates query-conditioned role specifications, builds a minimal execution DAG, and iteratively refines both roles and structure using lightweight feedback signals. Use when: - "Set up a multi-agent pipeline to solve this problem" - "Use self-evolving agents to generate and debug this code" - "Orchestrate multiple specialized agents for this reasoning task" - "Dynamically create roles to handle this complex task" - "Build an adaptive agent topology for this multi-step problem" - "Use MetaGen-style agent collaboration"

ndpvt-web Skill

Translate natural language optimization problems into mathematical models and solver code using MIRROR's multi-agent pipeline with iterative error correction and hierarchical retrieval. Use when: 'solve this optimization problem', 'write a linear program for', 'model this scheduling/routing/allocation problem', 'convert this OR problem to Gurobi code', 'formulate constraints for this optimization', 'help me model this integer program'.

ndpvt-web Skill

Implement a Planner-Auditor twin architecture that decouples LLM generation from deterministic validation with self-improvement loops. Use when: 'build a planner-auditor pipeline', 'add deterministic auditing to my LLM agent', 'implement self-improving generation with confidence calibration', 'create a two-tier feedback loop for LLM outputs', 'build a FHIR discharge planning system', 'add discrepancy buffering and replay to my agent'

ndpvt-web Skill

Systematically extract deep knowledge from LLMs using an interactive agentic framework with four adaptive exploration policies and a three-stage deduplication pipeline. Use this skill when the user says: "extract everything the model knows about X", "probe knowledge boundaries", "deep knowledge extraction", "exhaustive topic mining", "knowledge audit of an LLM", or "map out what you know about X".

ndpvt-web Skill

Build tree-structured, agent-assisted thought-to-video authoring systems where each generation step is a node binding intent, prompts, parameters, and outputs. Four collaborating agents (Master, Knowledge, Workflow, Prompt) translate user intent into editable executable plans. Supports branching exploration, provenance tracking, and convergent stitching assembly. Trigger phrases: "build a video authoring pipeline", "tree-based video generation", "agent-assisted video creation", "thought-to-video workflow", "branching video exploration system", "multi-scene video authoring with agents"

ndpvt-web Skill

Explain generative AI outputs using the gSMILE perturbation-based attribution framework. Builds local surrogate models from controlled input perturbations and Wasserstein distance to produce token-level or word-level importance scores for LLM and diffusion model outputs. Triggers: 'explain why the model generated this', 'token attribution for prompt', 'which words in my prompt matter most', 'interpret generative model output', 'build explainability for my LLM pipeline', 'debug prompt influence on generation'

ndpvt-web Skill

Detect and prevent semantic collapse in LLM outputs — where models fabricate spurious relationships between unrelated concepts. Apply CORE-style ontological relation evaluation to audit code, data pipelines, knowledge graphs, and AI systems for unrelatedness reasoning failures. Use when: 'check if these concepts are actually related', 'audit my ontology for spurious relations', 'evaluate semantic relationships in my knowledge graph', 'detect hallucinated connections in LLM output', 'validate entity relationships in my schema', 'test unrelatedness reasoning in my AI pipeline'.

ndpvt-web Skill

Implement hierarchical meta-cognitive monitoring for deep search agents. Embeds a two-tier self-monitoring system (fast consistency checks + slow experience-driven reflection) into multi-step retrieval-reasoning loops to detect anomalies, prevent reasoning drift, and trigger corrective interventions. Use when: 'build a deep search agent with self-monitoring', 'add metacognitive monitoring to my search pipeline', 'detect and fix reasoning failures in multi-step retrieval', 'implement DS-MCM for search quality', 'add anomaly detection to my RAG agent', 'build a self-correcting research agent'.

ndpvt-web Skill

Build dialect-aware RAG conversational agents that handle non-standard orthography, code-switching, and multi-script input. Uses a dual-path architecture: deterministic NLU for structured flows + RAG fallback for open-domain queries. Trigger phrases: 'build a dialect chatbot', 'RAG agent for Arabic dialect', 'handle code-switching in chatbot', 'multi-script NLU pipeline', 'Algerian Arabic conversational agent', 'dialect-aware customer service bot'

ndpvt-web Skill

Build TabRAG-style pipelines that retrieve relevant tables from large image collections and answer natural language queries over them using multimodal LLMs. Implements a three-stage retrieve-rerank-reason architecture for table question answering at scale. Trigger phrases: - "find the right table and answer my question" - "search across table images to answer a query" - "build a table retrieval pipeline" - "RAG over table images" - "table QA from document scans" - "retrieve and reason over tabular data"

ndpvt-web Skill

Design and optimize Kubernetes-native GenAI inference platforms using Kueue job queuing, Dynamic Accelerator Slicer (DAS) GPU partitioning, and Gateway API Inference Extension (GAIE) with llm-d for multi-stage AI pipelines. Use when: 'set up Kubernetes for AI inference', 'configure Kueue for batch GPU jobs', 'partition GPUs with MIG slicing on Kubernetes', 'optimize LLM inference routing on Kubernetes', 'build a Whisper-to-LLM pipeline on K8s', 'reduce TTFT latency for LLM serving'.

ndpvt-web Skill

Build intelligent trigger-action automation systems using FARM's two-stage architecture: contrastive retrieval + multi-agent LLM selection with field-level configuration. Use when asked to 'create an IFTTT-style automation', 'build a trigger-action pipeline', 'connect event triggers to actions with field bindings', 'design a multi-agent workflow resolver', 'generate executable automation rules from natural language', or 'wire up webhook triggers to API actions'.

ndpvt-web Skill

Analyze SEC filings and financial disclosures using the Fin-RATE three-pathway methodology: detail-oriented reasoning within single documents, cross-entity comparison across companies, and longitudinal tracking across reporting periods. Includes structured error diagnosis for retrieval, generation, reasoning, and context failures. Use when: 'analyze this 10-K filing', 'compare revenue across these companies', 'track this firm's risk factors over time', 'build a financial QA pipeline over SEC filings', 'evaluate RAG accuracy on regulatory documents', 'diagnose why my financial QA system hallucinates'.

ndpvt-web Skill

Harden ICL classification prompts against zero-query black-box evasion attacks. Audit in-context learning pipelines for Fake Claim, Template, and Needle-in-a-Haystack vulnerabilities, then apply the joint defense recipe. Triggers: 'harden my ICL prompt', 'audit ICL classifier security', 'defend against prompt evasion attacks', 'ICL adversarial robustness', 'protect few-shot classifier from manipulation', 'red-team my in-context learning pipeline'

ndpvt-web Skill

Build rigorous, multi-task evaluation benchmarks for domain-specific LLMs using the LingLanMiDian methodology: synonym-tolerant matching, difficulty-ranked hard subsets, character-level F1, and decision recognition reframing. Trigger phrases: 'evaluate LLM on domain knowledge', 'build a medical benchmark', 'TCM evaluation pipeline', 'synonym-tolerant scoring', 'create hard subset for benchmark', 'domain-specific LLM evaluation'

ndpvt-web Skill

Automated Docker environment construction for polyglot repositories using a Planning-Execution-Verification multi-agent loop with environment reuse. Use when: 'build a Docker environment for this repo', 'set up a reproducible test environment', 'create a verifiable dev container', 'construct an executable environment for this project', 'make this repo's tests runnable in Docker', 'set up CI environments for multiple languages'.

ndpvt-web Skill

Build polarized memory systems for multimodal agents that encode both positive and negative evidence as graph constraints, suppressing hallucinations without retraining. Use when: 'add negative evidence to RAG memory', 'build verifiable retrieval for VLM agent', 'suppress hallucinations in multimodal retrieval', 'polarized graph memory for agent', 'encode negation constraints in memory', 'logic-dominant retrieval pipeline'.

ndpvt-web Skill

Build agentic pipelines that ingest heterogeneous raw scientific data, parse research intent, and produce analysis-ready unified datasets. Use when user says 'prepare scientific data', 'build a data ingestion pipeline', 'normalize heterogeneous datasets', 'parse raw experimental data', 'integrate multi-modal scientific data', or 'make this data AI-ready'.

ndpvt-web Skill

Deploy SERE (Similarity-based Expert Re-routing) to accelerate MoE model batch decoding in vLLM by dynamically skipping redundant experts. Use when: 'speed up MoE inference', 'optimize Qwen MoE serving', 'reduce MoE expert activation overhead', 'SERE expert re-routing', 'batch decoding latency MoE', 'vLLM MoE throughput optimization'

ndpvt-web Skill

Defend vision-language models (VLMs) against universal and transferable adversarial image attacks using techniques from UltraBreak (ICLR 2026). Helps build robust VLM pipelines by implementing adversarial robustness evaluations, input sanitization, and detection mechanisms grounded in the vision-space regularisation and semantic loss landscape insights from Cui et al. Trigger phrases: - "harden my VLM against adversarial images" - "evaluate VLM robustness to image-based jailbreaks" - "add adversarial image detection to my multimodal pipeline" - "build a red-team evaluation for my vision-language model" - "implement input sanitization for VLM image inputs" - "test if my VLM is vulnerable to transfer attacks"

ndpvt-web Skill

Build multi-agent adaptive learning systems that diagnose knowledge gaps and recommend targeted resources. Implements the ALIGNAgent framework: Skill Gap Agent (proficiency estimation + concept-level diagnostic reasoning) and Recommender Agent (preference-aware resource retrieval aligned to deficiencies). Trigger phrases: - "Build an adaptive learning system" - "Create a personalized tutoring agent" - "Diagnose student knowledge gaps from quiz data" - "Build a skill gap analyzer for learners" - "Create an educational recommender that adapts to student performance" - "Implement a multi-agent pipeline for personalized education"

ndpvt-web Skill

Dynamically create specialized sub-agents for complex multi-step tasks using the AOrchestra pattern: decompose goals, then spawn tailored (Instruction, Context, Tools, Model) executors on-the-fly. Use when: 'break this task into sub-agents', 'orchestrate agents for this problem', 'create a multi-agent workflow', 'delegate subtasks to specialized agents', 'build an agent pipeline for this', 'dynamically assign agents to subtasks'.

ndpvt-web Skill

Generate high-reliability test suites without ground-truth implementations using the ConVerTest pipeline: Self-Consistency voting, Chain-of-Verification refinement, and Dual Execution Agreement. Use when asked to 'generate tests for this spec', 'write tests before implementation', 'create a test suite without reference code', 'test-driven development for this feature', 'generate reliable unit tests', or 'validate tests without a working implementation'.

ndpvt-web Skill

Guides cost-aware model selection for text classification pipelines, applying multi-objective trade-off analysis (F1 vs cost vs latency) to choose between fine-tuned encoders (BERT/RoBERTa/DistilBERT) and LLM prompting (GPT-4o/Claude). Uses Pareto frontier analysis and a parameterized utility function to recommend the right model for a given deployment regime. Trigger phrases: - "Which model should I use for text classification?" - "Is GPT-4o overkill for my classification task?" - "Help me pick a cost-effective NLP model" - "Compare BERT vs LLM for classification cost" - "Optimize my text classification pipeline for production" - "Build a cost-aware NLP system"

ndpvt-web Skill

Evaluate and select LLMs using CreditAudit's 2D framework: mean ability plus stability risk (fluctuation) across system prompt variations. Assigns credit grades (AAA–BBB) to models based on performance volatility. Use when: 'compare models for deployment', 'which LLM is most stable', 'evaluate model robustness to prompt changes', 'credit grade these models', 'model selection for agentic pipeline', 'rank models by reliability'.

ndpvt-web Skill

Build multi-agent pipelines that transform CVE metadata into fully executable vulnerability reproduction environments with Docker, automated tests, and verified patches. Use this skill when: - "Set up a CVE reproduction environment" - "Create an executable security task from a CVE" - "Build a vulnerability benchmark with Docker" - "Reproduce CVE-2025-XXXXX in an isolated container" - "Generate exploit tests and patch verification for a vulnerability" - "Design a multi-agent pipeline for security task automation"

ndpvt-web Skill

Constrain LLM generation with executable Pydantic schemas and multi-agent pipelines to produce structurally valid, domain-rich artifacts. Uses ontology-as-grammar to eliminate hallucinated structures while preserving creative output. Trigger phrases: "generate a valid game design", "schema-constrained generation", "build a multi-agent pipeline with Pydantic validation", "ontology-driven content generation", "structured creative generation with DSPy", "generate artifacts that pass domain validation".

ndpvt-web Skill

Build neuro-symbolic safety verification pipelines using the VIRF (Verifiable Iterative Refinement Framework) pattern: a Logic Tutor provides formal, causal feedback to an LLM planner, enabling intelligent plan repair instead of mere rejection. Use this skill when: - "Add safety verification to my LLM agent pipeline" - "Build a plan validator with formal logic feedback" - "Create a tutor-apprentice loop for safe AI planning" - "Implement iterative plan refinement with ontology checks" - "Add OWL-based safety constraints to my agent" - "Build a verifiable planning system with correction feedback"

ndpvt-web Skill

Evaluate LLM-powered recommender systems across five human-centered dimensions: Intent Alignment, Explanation Quality, Interaction Naturalness, Trust & Transparency, and Fairness & Diversity. Use when: 'evaluate my recommendation system', 'audit my LLM recommender for bias', 'score my chatbot recommendations', 'measure explanation quality of my recommender', 'check fairness of my recommendation engine', 'run a human-centered evaluation on my rec system'.

ndpvt-web Skill

Build conversational Earth Observation agents that turn natural-language queries into executable, auditable Python workflows. Uses a unified API covering classification, segmentation, detection, spectral indices, and geospatial operations. Trigger phrases: 'analyze satellite imagery', 'earth observation pipeline', 'land cover classification code', 'post-wildfire damage assessment', 'generate EO workflow', 'spectral index calculation'.

ndpvt-web Skill

Evaluate the logical integrity of LLM-generated multi-hop answers using Horn Rule backward chaining. Scores Completeness (gap-free reasoning), Conciseness (no redundant steps), and Determinateness (answer entailment). Use when: 'evaluate my QA pipeline logic', 'check reasoning chain completeness', 'score answer conciseness', 'find deductive gaps in generated answers', 'audit multi-hop reasoning quality', 'LogicScore evaluation'.

ndpvt-web Skill

Performance-aware multi-tool orchestration for visual content generation pipelines. Implements PerfGuard's three mechanisms (PASM, APU, CAPO) to select, score, and schedule AI image/video tools based on measured capability boundaries instead of generic descriptions. Use when: "build a visual generation pipeline", "orchestrate multiple image tools", "select the best AI model for this image task", "score and rank generation tools", "adaptive tool selection for AIGC", "performance-aware tool routing".

ndpvt-web Skill

Build lightweight hallucination detection pipelines using Predictive Coding surprise signals and Information Bottleneck perturbation testing. Implements the PCIB framework: extract interpretable signals (Uptake, Stress, Conflict, Falsifiability) from LLM outputs, train sub-1M-parameter classifiers, and deploy real-time hallucination filters for RAG systems. Trigger phrases: "detect hallucinations", "hallucination detection pipeline", "verify LLM output", "build a hallucination filter", "RAG output quality scoring", "factual grounding check"