DevOps & Infra

ndpvt-web Skill

Audit and instrument agentic AI systems for system-level interpretability and accountability. Embeds traceability, causal analysis, and oversight mechanisms across the agent lifecycle—from goal formation through environmental interaction to outcome evaluation. Use when: 'add observability to my agent pipeline', 'trace why my agent made this decision', 'audit my multi-agent system', 'add interpretability logging to my LLM agent', 'debug compounding errors in my agent chain', 'instrument my agentic workflow for accountability'.

ndpvt-web Skill

Extract and normalize skills from job postings using a two-stage LLM pipeline: dynamic few-shot skill identification followed by embedding retrieval + LLM reranking against a standardized taxonomy (ESCO). Optimized for morphologically complex and low-resource languages like Turkish. Use when: 'extract skills from job descriptions', 'parse job postings for competencies', 'match skills to ESCO taxonomy', 'build a skill extraction pipeline', 'normalize skills from Turkish job ads', 'link extracted skills to a standard taxonomy'.

ndpvt-web Skill

Classify and diagnose intermittent CI/CD job failures from execution logs using the FlaXifyer few-shot approach and LogSift log reduction. Trigger phrases: 'diagnose flaky build', 'classify CI failure', 'why did this pipeline fail intermittently', 'triage job log', 'find root cause in CI log', 'reduce CI log noise'

ndpvt-web Skill

Audit fine-tuning datasets and pipelines for privacy collapse — the silent failure where benign training data degrades a model's contextual privacy reasoning while safety benchmarks stay green. Use when: 'audit my fine-tuning data for privacy risks', 'check if this dataset causes privacy collapse', 'evaluate privacy after fine-tuning', 'scan training data for privacy-degrading patterns', 'test model for contextual privacy norms', 'build a privacy-safe fine-tuning pipeline'.

ndpvt-web Skill

Analyze and optimize token consumption in LLM-based multi-agent software engineering workflows. Maps agent execution traces to SDLC stages (Design, Coding, Code Completion, Code Review, Testing, Documentation) and quantifies where tokens are spent. Use when: 'analyze token usage in my agent pipeline', 'where are tokens being wasted in my workflow', 'optimize my multi-agent system costs', 'profile my agentic coding pipeline', 'reduce LLM costs in my CI/CD agents', 'tokenomics analysis of my agent framework'.

ndpvt-web Skill

Red-team LLM-based agentic payment systems against prompt injection attacks targeting transaction integrity and credential isolation. Use when: 'red-team my payment agent', 'test agent payment security', 'prompt injection audit for shopping agent', 'secure my AP2 implementation', 'harden agentic checkout flow', 'find vulnerabilities in my LLM agent pipeline'.

ndpvt-web Skill

Select and configure the right RAG strategy for conversational QA systems based on dataset characteristics. Use when: 'build a conversational RAG pipeline', 'choose a RAG method for multi-turn QA', 'my RAG pipeline performs worse than no retrieval', 'optimize retrieval for dialogue systems', 'compare RAG strategies for my dataset', 'reranking vs HyDE vs hybrid BM25'.

ndpvt-web Skill

Build multi-agent workflows structured as constrained DAG process maps with Monte Carlo uncertainty estimation. Each agent occupies a specialized review role, with predefined escalation paths and terminal states (automated label or human review). Use when: 'build a multi-agent compliance pipeline', 'add uncertainty-aware escalation to my agent workflow', 'create a review chain with human fallback', 'design a DAG-based agent workflow with confidence thresholds', 'implement Monte Carlo sampling for agent decisions', 'build a multi-stage content moderation system'.

ndpvt-web Skill

Detect and fix cross-lingual evaluation instabilities in LLM-as-a-judge pipelines. Use when: 'audit my multilingual eval pipeline', 'check if my LLM judge is stable across languages', 'set up cross-lingual evaluation', 'calibrate judge scoring for non-English languages', 'diagnose ranking inversions in multilingual benchmarks', 'build controlled generation tests for eval reliability'.

ndpvt-web Skill

Implements a curiosity-driven knowledge retrieval framework for autonomous agents. Formalizes agent uncertainty as a curiosity score, triggers external knowledge retrieval when uncertainty exceeds a threshold, and organizes retrieved knowledge into structured AppCards for selective integration into reasoning. Trigger phrases: "build an agent with curiosity-driven retrieval", "add uncertainty-aware knowledge lookup", "implement AppCard knowledge system", "create a curiosity-scored agent pipeline", "build adaptive knowledge retrieval", "implement uncertainty-triggered documentation lookup"

ndpvt-web Skill

Evolutionary multi-agent code optimization using genetic algorithms. Agents mutate each other's training/configuration code, benchmark results, and select survivors across generations. Use when: 'evolve my training config', 'optimize this code with genetic search', 'set up evolutionary hyperparameter tuning', 'multi-agent code mutation pipeline', 'self-improving training loop', 'darwin-style evolutionary optimization'.

ndpvt-web Skill

Extract security-relevant signals from code review comments and translate them into fuzzer-guiding annotations using the EyeQ pipeline. Use when the user says 'guide fuzzing from code reviews', 'find dragons in review comments', 'annotate code for fuzzing', 'review-guided fuzzing', 'extract security signals from PRs', or 'instrument code for AFL++ from review discussions'.

ndpvt-web Skill

Diagnose and explain failures in agentic AI systems using trace-based rubric evaluation, bridging static feature attribution (SHAP/LIME) with trajectory-level diagnostics. Use when: 'debug why my agent failed', 'explain agent behavior', 'evaluate agent traces', 'add explainability to my agent pipeline', 'diagnose agentic failures', 'trace-based agent analysis'.

ndpvt-web Skill

Build hierarchical causal knowledge graphs for RAG pipelines that suppress spurious correlations and enable cross-document causal reasoning. Use when: 'build a causal knowledge graph from my documents', 'add causal reasoning to my RAG pipeline', 'set up graph-based RAG with causal filtering', 'create a hierarchical knowledge graph for retrieval', 'implement causal gating for my knowledge base', 'improve RAG with structured causal retrieval'.

ndpvt-web Skill

Build immutable, agent-native medical data pipelines using Merkle DAG structures (MedBeads pattern). Converts mutable EMR/FHIR records into cryptographically-linked, causally-ordered bead graphs that LLMs can traverse deterministically instead of relying on probabilistic RAG. Use when: 'build an immutable clinical data store', 'convert FHIR to a causal DAG', 'create tamper-evident medical records', 'agent-native healthcare data pipeline', 'deterministic context retrieval for medical AI', 'MedBeads implementation'.

ndpvt-web Skill

Build iterative search-and-reason agents for scientific literature QA. Uses the PaperSearchQA pattern: interleaved thinking, querying, and verification loops over document corpora. Trigger phrases: 'search scientific papers', 'build a paper search agent', 'scientific QA pipeline', 'iterative retrieval agent', 'RLVR search agent', 'reason over research papers'

ndpvt-web Skill

Build privacy-aware pipelines that filter sensitive content from visual frames before sending to cloud AI models, using edge preprocessing with object detection (YOLO), text-based scene description, selective cropping, and structured MLLM interaction. Triggers: 'privacy-aware XR pipeline', 'filter sensitive data from camera frames', 'edge preprocessing before cloud AI', 'PRISM-XR privacy pipeline', 'sanitize visual input for LLM', 'multi-user XR collaboration with privacy'

ndpvt-web Skill

Automated feature engineering for prediction tasks on relational databases using a multi-agent LLM pipeline. Generates, filters, and validates SQL/pandas features from multi-table schemas. Use when: 'generate features from my database for prediction', 'engineer features across related tables', 'automate feature generation for classification/regression on relational data', 'build predictive features from a star schema', 'improve model accuracy with relational feature engineering', 'create aggregation features from joined tables'.

ndpvt-web Skill

Optimize Pandas code by discovering per-program improvements, generalizing them into reusable rewrite rules, and applying those rules as a lightweight compiler pass. Use when the user says 'optimize this Pandas code', 'speed up my DataFrame operations', 'rewrite my notebook for performance', 'apply Pandas rewrite rules', 'find slow Pandas patterns', or 'generate optimization rules for my Python data pipeline'.

ndpvt-web Skill

Automated game balancing using multi-agent LLM self-play coupled with Bayesian optimization. Use when the user asks to 'balance a game', 'tune game parameters', 'optimize game rules', 'automate playtesting', 'build a game balancing pipeline', or 'use LLM agents for game simulation'.

ndpvt-web Skill

Optimize deep multi-step AI pipelines using Textual Equilibrium Propagation (TEP) — a two-phase local-then-nudge strategy that avoids gradient explosion/vanishing in long prompt chains. Use when: "optimize my multi-agent pipeline", "fix my prompt chain that degrades at depth", "improve my compound AI system", "TEP optimization", "local critic refinement for agents", "scale my LLM workflow without losing quality".

ndpvt-web Skill

Detect, measure, and defend against collective false-memory propagation (the Mandela Effect) in LLM multi-agent systems. Use when: 'harden multi-agent pipeline against misinformation', 'audit agent consensus for false beliefs', 'add cognitive anchoring to agent prompts', 'defend agents from social influence bias', 'reduce reality shift in collaborative agents', 'mitigate groupthink in LLM swarms'.

ndpvt-web Skill

Extract and classify product attributes from images using Vision-Language Models with structured prompts and a three-tier evaluation framework. Handles conditional attributes (e.g., fabric type only when garment is visible) by separating applicability detection from classification. Triggers: 'extract product attributes from images', 'label fashion attributes zero-shot', 'classify clothing attributes with VLM', 'build product catalog enrichment pipeline', 'detect garment properties from photos', 'evaluate attribute prediction accuracy'.

ndpvt-web Skill

Build AI-augmented annotation pipelines for creating high-quality information retrieval and QA datasets. Combines LLM-generated suggestions (questions, passage relevance scores, answer spans) with human review workflows to accelerate dataset creation. Use when: 'build an annotation pipeline for IR data', 'create a QA dataset from documents', 'annotate passages for retrieval training', 'generate question-answer pairs from a corpus', 'build a human-in-the-loop labeling tool for search', 'set up AI-assisted relevance annotation'.

ndpvt-web Skill

Generate fine-grained evaluation rubrics for medical dialogue systems using a retrieval-augmented multi-agent pipeline. Decomposes medical evidence into atomic facts, synthesizes them with interaction constraints, and produces weighted, auditable rubrics. Use when: 'evaluate medical chatbot responses', 'generate rubrics for clinical QA', 'build a medical LLM evaluation pipeline', 'score health dialogue quality', 'create automated clinical evaluation criteria', 'refine medical AI responses with rubric feedback'.

ndpvt-web Skill

Evaluate and defend RAG-based code generation systems against energy-drain attacks that poison retrieval contexts to inflate LLM output length, latency, and GPU energy consumption. Use when: 'audit my RAG pipeline for energy attacks', 'test code retrieval poisoning resilience', 'detect adversarial triggers in retrieved code', 'harden my code generation system against context poisoning', 'benchmark energy cost of retrieval-augmented code generation', 'simulate DrainCode-style attacks on my pipeline'.

ndpvt-web Skill

Detect and defend against fraudulent content in LLM inputs using knowledge-graph-augmented analysis. Builds a fraud tactic-keyword bipartite graph, scores associations by confidence, prunes ambiguities, and augments prompts with XML-tagged keywords plus evidence rationales. Use when: 'check this email for fraud', 'is this job posting a scam', 'analyze this contract for manipulation', 'detect phishing in this message', 'flag suspicious text in this document', 'add fraud detection to my LLM pipeline'.

ndpvt-web Skill

Generate cross-domain agentic workflows using decompose-recompose-decide composition over reusable capability bases. Use when asked to: 'design a multi-step agent workflow', 'create a pipeline that works across different task types', 'build a reusable agentic workflow', 'compose agent operators for a new domain', 'generate a single-pass workflow for complex tasks', 'orchestrate reasoning-verification-repair agents'.

ndpvt-web Skill

Implements a reflective dual-agent (Annotator + Reviewer) workflow for automated linguistic annotation tasks such as metaphor identification, sentiment labeling, named entity recognition, and other sequence-labeling problems. The Annotator marks spans in text using XML tags and provides reasoning; the Reviewer critiques the annotations against a codebook, catching false positives and missed instances, then feeds corrections back for self-improvement. Trigger phrases: "annotate this text for metaphors", "dual-agent annotation pipeline", "reflective annotation workflow", "linguistic annotation with LLM review", "peer-review annotation system", "automated text labeling with self-correction"

ndpvt-web Skill

Decouple visual perception from reasoning when building VLM pipelines, image analysis agents, or multi-modal workflows. Implements the SPARC two-stage pattern: first localize question-relevant regions via visual search, then condition reasoning only on those regions. Dramatically reduces token budgets and improves accuracy on visually complex tasks. Trigger phrases: - "Build an image QA pipeline that scales at test time" - "Analyze this high-resolution image efficiently with a VLM" - "Separate perception from reasoning in my vision pipeline" - "Reduce token cost for visual reasoning tasks" - "Implement a two-stage visual search and answer system" - "Scale VLM inference with asymmetric compute allocation"

ndpvt-web Skill

Build and evaluate domain-specific RAG systems with iterative user-feedback refinement, source grounding, and structured evaluation pipelines. Use when: 'build a RAG system for [domain] documents', 'evaluate my RAG pipeline quality', 'add user feedback to my retrieval system', 'improve RAG answers with source grounding', 'set up domain-specific document retrieval', 'create an evaluation framework for my RAG app'.

ndpvt-web Skill

Adaptive multi-step tool orchestration for complex reasoning tasks. Dynamically selects, sequences, and composes tools based on task context and intermediate results rather than fixed pipelines. Use when: 'orchestrate tools for this task', 'figure out which tools to use', 'multi-step reasoning with tools', 'adaptive tool pipeline', 'dynamic tool selection', 'chain tools together intelligently'.

ndpvt-web Skill

Reverse-engineer black-box agentic systems into editable, interpretable workflows using search-based reconstruction. Use when the user says 'reconstruct this agent workflow', 'reverse-engineer this pipeline', 'white-box this agentic system', 'explain what this agent chain is doing', 'approximate this black-box agent', or 'build an interpretable surrogate for this system'.

ndpvt-web Skill

Build self-managing data processing pipelines using hierarchical meta-agent orchestration. Decomposes complex data tasks into multi-phase plans, instantiates specialized ground-level agents (Reader, Profiler, Transformer, Validator, Joiner, etc.), and iteratively refines execution through progressive sampling and monitoring loops. Triggers: 'build a data pipeline', 'process and clean this dataset', 'create an ETL pipeline with agents', 'orchestrate data processing agents', 'autonomous data pipeline', 'meta-agent data processing'.

ndpvt-web Skill

Cross-cultural meme transcreation using a three-stage hybrid pipeline (cultural analysis, visual generation, assembly) that preserves humor and communicative intent while adapting culture-specific references between languages. Triggers: 'transcreate this meme', 'adapt meme for Chinese audience', 'convert meme to US culture', 'cross-cultural meme adaptation', 'localize this meme for another culture', 'meme cultural translation'

ndpvt-web Skill

Analyze and optimize multi-agent code generation pipelines using causality-based importance ranking of intermediate features. Identifies which pipeline stages matter most, enables targeted failure repair, token-efficient pruning, and hybrid LLM backend assignment. Triggers: 'analyze my multi-agent pipeline', 'optimize agent code generation', 'which pipeline stages matter most', 'reduce token usage in my agent system', 'fix failing multi-agent code generation', 'assign LLMs to pipeline stages'