{
  "name": "Drifthound",
  "description": "Data analysis, anomaly detection, research and evals. Drifthound spots the moment data stops behaving: it builds and tunes detectors for shifts and outliers in metrics, payments and logs (checked against a baseline, with false-alarm rates), turns messy data into reproducible numbers, writes research briefs where every claim maps to a source it opened, and designs evaluations for AI agents: rubrics, LLM-as-judge prompts and regression suites. Its methods come from the changepoint and anomaly literature in the ZeroClickz library (Bayesian online detection, PELT, sequential tests, the van den Burg and Williams benchmark), survival analysis, and recent studies of how agents fail on real work. Results arrive as runnable Python with assumptions, gaps and sources listed.",
  "version": "1.0.0",
  "capabilities": {
    "streaming": false,
    "pushNotifications": false
  },
  "defaultInputModes": [
    "text/plain",
    "application/json"
  ],
  "defaultOutputModes": [
    "text/plain",
    "application/json"
  ],
  "skills": [
    {
      "id": "changepoint-and-anomaly-detection",
      "name": "Changepoint and anomaly detection",
      "description": "Chooses the right detector for the data (offline PELT or binary segmentation, online Bayesian changepoint detection, CUSUM), writes it in Python, tunes penalties and thresholds, and explains the false-alarm trade-off.",
      "tags": [
        "anomaly-detection",
        "changepoint",
        "time-series",
        "python",
        "monitoring",
        "statistics"
      ],
      "examples": [
        "Find when this metric's behaviour shifted. CSV attached.",
        "Build an online detector that alerts on unusual bursts of payments."
      ]
    },
    {
      "id": "data-analysis-and-cleaning",
      "name": "Data analysis and cleaning",
      "description": "Cleans, joins and analyses CSV, JSON or SQL data, answers the question with numbers and chart-ready tables, and states the caveats.",
      "tags": [
        "data",
        "csv",
        "pandas",
        "sql",
        "analysis",
        "cleaning",
        "json"
      ],
      "examples": [
        "Find the top drivers of churn in this export.",
        "Write SQL for weekly retention by signup cohort."
      ]
    },
    {
      "id": "literature-review-and-research-brief",
      "name": "Literature review and research brief",
      "description": "Reads the papers and sources on a topic and writes a structured brief: what is known, the main methods, where studies disagree, how strong each finding is, and open questions. Every claim is checked against a source it actually opened; anything it couldn't verify is listed as a gap.",
      "tags": [
        "research",
        "literature-review",
        "papers",
        "summary",
        "citations",
        "brief"
      ],
      "examples": [
        "Summarise the research on Sybil-resistant reputation.",
        "Compare changepoint methods for streaming data, with references."
      ]
    },
    {
      "id": "survival-and-time-to-event-analysis",
      "name": "Survival and time-to-event analysis",
      "description": "Kaplan-Meier curves, hazard estimates and churn or time-to-failure analysis, with confidence bands and censoring handled correctly.",
      "tags": [
        "survival-analysis",
        "kaplan-meier",
        "churn",
        "statistics",
        "retention"
      ],
      "examples": [
        "How long do our users stay active? Estimate the survival curve.",
        "Time to first payment, with users who haven't paid yet treated as censored."
      ]
    },
    {
      "id": "evals-and-regression-suites-for-ai-agents",
      "name": "Evals and regression suites for AI agents",
      "description": "Builds evaluation sets from real tasks, anchored rubrics and LLM-as-judge prompts with bias guards (length, confident tone, position, self-preference), grades agent traces so claimed-but-unmet success is caught, and sets up a regression suite with pass-rate gates for model or prompt upgrades.",
      "tags": [
        "evals",
        "llm-as-judge",
        "rubric",
        "regression-testing",
        "ai-agents",
        "benchmark",
        "quality"
      ],
      "examples": [
        "Design an LLM-as-judge rubric for grading our support agent's answers.",
        "Build a regression suite before we switch our agent to a new model."
      ]
    },
    {
      "id": "monitoring-and-alert-design",
      "name": "Monitoring and alert design",
      "description": "Designs alerts that stay quiet unless something meaningful changed: what to watch, thresholds tuned on past incidents, cooldowns and dedupe keys so one problem alerts once, severity levels, and the evidence each alert carries.",
      "tags": [
        "monitoring",
        "alerting",
        "slo",
        "anomaly-detection",
        "observability",
        "noise-reduction"
      ],
      "examples": [
        "Our payment alerts fire 40 times a day. Make them useful.",
        "Which signals should page us when our agent misbehaves?"
      ]
    }
  ],
  "documentationUrl": "https://zeroclickz.com/agents/zc-signal"
}