ai-problem-packages

# AI Problem Packages Zero-dependency npm packages for common AI engineering problems: - `@mukundakatta/prompt-injection-shield`: scan untrusted text for prompt-injection patterns. - `@mukundakatta/pii-sentry`: detect and redact common PII and secret-like values. - `@mukundakatta/rag-quality-kit`: score retrieval coverage, answer grounding, and citation coverage. - `@mukundakatta/model-router-policy`: route requests by capability, cost, latency, and privacy policy. - `@mukundakatta/llm-cost-guard`: estimate token usage and enforce request/session budgets. - `@mukundakatta/tool-call-contracts`: validate tool-call payloads against small JSON-like schemas. - `@mukundakatta/eval-dataset-smith`: build balanced eval datasets from bugs, docs, and examples. - `@mukundakatta/agent-trajectory-replay`: replay and diff agent event trajectories. - `@mukundakatta/context-drift-detector`: detect topic drift between prompt, retrieved context, and answer. - `@mukundakatta/semantic-cache-key`: create normalized semantic cache keys for AI requests. - `@mukundakatta/agent-regression-lens`: detect regressions between baseline and current AI agent runs. - `@mukundakatta/llm-output-sanitizer`: Sanitize LLM outputs before rendering, SQL, shell, or markdown sinks. - `@mukundakatta/tool-permission-gate`: Policy-check agent tool calls before execution. - `@mukundakatta/system-prompt-leak-scan`: Detect system prompt leakage in model outputs. - `@mukundakatta/vector-poison-score`: Score retrieved documents for vector/RAG poisoning signals. - `@mukundakatta/retrieval-acl-filter`: Enforce document ACLs after retrieval and before prompting. - `@mukundakatta/citation-integrity-check`: Verify answer citations refer to supplied source ids. - `@mukundakatta/hallucination-risk-meter`: Estimate hallucination risk from answer, context, citations, and uncertainty language. - `@mukundakatta/prompt-token-trim`: Trim prompt messages to fit a token budget while preserving priority. - `@mukundakatta/eval-flake-detector`: Detect flaky LLM eval cases across repeated runs. - `@mukundakatta/llm-response-schema-lite`: Tiny schema validator for structured LLM responses. - `@mukundakatta/agent-loop-breaker`: Detect repeated agent steps and stop runaway loops. - `@mukundakatta/tool-result-taint`: Track untrusted tool output before it enters prompts or actions. - `@mukundakatta/rag-staleness-auditor`: Find stale RAG chunks by age, version, and freshness requirements. - `@mukundakatta/model-fallback-planner`: Plan model fallback chains from capability, cost, and health data. - `@mukundakatta/prompt-version-diff`: Diff prompt templates and flag risky instruction changes. - `@mukundakatta/consent-redaction-log`: Record consent-aware redactions for privacy review trails. - `@mukundakatta/embedding-dedupe`: Deduplicate near-identical embedding records by cosine similarity. - `@mukundakatta/context-window-packer`: Pack context chunks into a budget by relevance and priority. - `@mukundakatta/jailbreak-corpus-mini`: Small local jailbreak and prompt-injection fixture set for tests. - `@mukundakatta/ai-supply-chain-manifest`: Build and validate lightweight AI model/data/tool manifests. - `@mukundakatta/llm-trace-sampler`: Sample LLM traces by risk, errors, latency, and deterministic ids. - `@mukundakatta/openai-responses-testkit`: Test OpenAI Responses API outputs, streams, and tool calls with snapshot-safe helpers.