Audits | Dagnam.AI Docs

Run a workload audit from the CLI, then read, share, and act on its report in the web app.

Audits An audit reads your own LLM traces, groups them into workloads, and reports which ones are worth replacing with a small model you own. This page covers the web app side: the audits list, reading a report, and acting on it. The audit itself runs from the CLI. What an audit tells you "What your LLM workloads cost, and what a replacement would." Each workload gets a verdict, and the workloads worth replacing get a candidate model, trained and scored against your own held-out traffic. Run one Audits start from the dagnam CLI on your own machine; see Workload audit /docs/dag-lib/workload-audit for the full walkthrough. Once you run it, the audit also appears here, unless you choose to keep it local. Training and serving candidates with dagnam audit run needs a personal API key, which is not on the Free plan, and cloud GPU access, which is available on request during the beta. Ask through Support /support . The audits list An empty account shows No audits yet with a link to install the CLI and scan. Otherwise, each row shows: Audit , Source , Status , Workloads audited of found , Best verdict , Credits , Spend / month , Projected savings , and Started . An audit's status is running, halted, or completed. Verdicts Verdict Meaning ------- ------------------------------------------------------------------- REPLACE A candidate cleared the quality bar and the economics favor it. NOT YET Worth replacing on economics, but no candidate cleared the bar yet. KEEP Not a good fit for a small model today. A workload's scan carries one of: Worth replacing , Marginal saving , Not worth replacing , Not audited , Too few samples , or Cost unknown . Reading a report The summary shows Spend today , Projected savings , Workloads audited and Ready to replace . Each workload lists its candidates, trained through the same pipeline: Upload , Split , PII check , Train , Deploy , Replay , Score . A candidate is a Head tune , SFT small , Hosted floor , or Zero shot model, and its card shows Agreement , Latency p50 / p95 , Serving / month and Credits . A collapsed Open models for comparison not run section lists reference models that were not run against your traffic. Switch to a winner When a workload has a winner, its Switch to this deployment card explains it: an OpenAI-compatible endpoint. Point your existing client at it; only the base URL, the model name and the key change. See OpenAI-compatible endpoint /docs/deployments/openai-compatible . The card also shows a rolling count of requests served and a Rotate key action. Retrain a workload Retrain queues one more candidate: choose a recipe Head tune text classification or QLoRA SFT chat , a base model, and optionally epochs and a learning rate, then Queue retrain . Share and download Share creates a read-only link to the report and copies it. Right after sharing, Revoke share link in the more actions menu turns it off; that item is offered only until you leave or reload the page. A revoked or expired link reads "This link is no longer available." Download exports the report as JSON or Markdown. Cancel and delete Cancel audit stops a running audit. Delete audit removes the audit and everything it created; it cannot be undone. Privacy Traces are redacted before anything leaves your machine, and the platform re-scans what it receives; a report carries counts and a masked template excerpt, never a raw prompt, and deployment keys live in your keyring rather than in anything shown here. Nothing is deleted on a schedule: the datasets, models and deployments an audit creates stay in your account until you delete the audit. See Privacy commitments /docs/dag-lib/workload-audit privacy-commitments .