TAO DEFT AOI Improvement Loop
Automates evaluation, defect augmentation, mining, retraining, and deployment gating for PCB inspection models.
The documentation provides a single pre-flight approval gate, path scoping, hard-stop rules, leakage checks, and atomic report writes, and explains the intended use of HF_TOKEN and NGC_KEY. However, it permits Bash, Write, Docker, autonomous training/data-generation/deployment gating, and asks users to enable auto-mode to bypass permission prompts. External model downloads, credential boundaries, deployment rollback, and recovery are incompletely specified, so points are deducted.
The workflow specifies state files, log sequence invariants, disk-as-source-of-truth behavior, blocker recovery, and hard stops in considerable detail. However, the supplied material lacks independently reproducible key-path tests; it depends on other skills and runtime helpers, and abnormal-input coverage remains limited. The static ceiling therefore requires a conservative deduction.
Triggers, supported model scope, non-fit cases, data contracts, and the required max_iterations input are clear. The skill nevertheless depends heavily on NVIDIA TAO, Docker, GPUs, HuggingFace, and related skills, gives no evidence of Chinese-language interaction or mainland-China network reachability, and requires additional onboarding when no workspace exists, so points are deducted.
The documentation uses layered references and includes licensing, authorship, version, script usage, data layout, report protocol, and limitations. Maintenance ownership, changelog, update path, and complete examples/FAQs are incomplete. The BENCHMARK also records missing recommended Instructions/Examples and schema/layout findings, so points are deducted.
The stated goal is to automate AOI baseline evaluation, RCA, synthetic defects, mining, retraining, and KPI gating, with state, logs, HTML reporting, and an inference specification as outputs. The BENCHMARK reports one external evaluation with an overall pass, but it covers only one task, provides no independent key-result reproduction, and depends on substantial user data and external components. Static calibration therefore limits this to 7.
The materials provide a pinned revision, detailed input/output contracts, an evaluation report, and auditable workflow rules. They do not provide multiple independent reproductions, comprehensive test coverage, or corroboration from multiple evidence types; the benchmark has only one task, so verification remains limited.
- Execution accesses external HuggingFace/NGC resources and uses sensitive credentials; confirm credential scope, network compliance, and cache locations before running.
- With auto-mode enabled, training, writes under the results directory, and deployment gating proceed continuously after pre-flight; confirm backups, target paths, and acceptable external effects.
- No evidence is provided for Chinese-language support or mainland-China network reachability, and the workflow depends on Docker, NVIDIA container tooling, GPUs, and several related skills/references not fully included here.
- The BENCHMARK contains only one evaluation task and cannot establish performance across datasets, hardware, or abnormal inputs.
What does this skill do, and when should you use it?
This skill runs the full DEFT AOI improvement loop for NVIDIA TAO VisualChangeNet/ChangeNet PCB inspection models. It covers baseline evaluation, root-cause analysis, Cosmos AnomalyGen/AMP synthetic defects, k-NN mining, retraining, and deployment gating. The loop continues until a FAR or recall KPI target is met, the iteration limit is reached, or a hard-stop condition fires. It requires Docker, nvidia-container-toolkit, workflow inputs, and the required credentials.
Reads workspace specifications, CSVs, datasets, checkpoints, and state files; performs pre-flight checks, baseline training or checkpoint recovery, inference, evaluation, and RCA; invokes related VisualChangeNet training, synthetic-defect, routing, and AOI mining workflows; updates deft_state.json and loop_log.jsonl during iterations; renders DEFT_Loop_Report.html after completed iterations and at loop end; and prepares an inference specification when a valid checkpoint exists. It presents one Pre-Flight Summary and waits for explicit approval before launching side-effecting work, then proceeds autonomously.
- An AOI engineer needs to reduce a ChangeNet model's false accept rate while preserving a target recall.
- A PCB inspection team needs RCA and synthetic defects to address model performance gaps.
- A machine-learning engineer wants repeated mining, retraining, and KPI checks starting from a baseline evaluation.
- A deployment team needs inference preparation gated on meeting FAR and recall thresholds.
What are this skill's strengths and limitations?
- Provides an end-to-end loop from baseline evaluation through deployment gating.
- Combines RCA, synthetic defects, k-NN mining, and iterative retraining.
- Uses a single pre-flight approval gate and explicit hard-stop conditions.
- Maintains state logs and HTML reports for iteration tracking.
- Scoped specifically to NVIDIA TAO VisualChangeNet/ChangeNet PCB inspection improvement.
- Not intended for standalone training, one-off inference, generic anomaly generation, or RCA-only analysis.
- Requires Docker, nvidia-container-toolkit, GPU workflow inputs, checkpoints, data, and credentials.
- The source does not state pricing or quotas for HF or NGC-related access.
- No standalone test suite or concrete platform validation results are described in the supplied source.
- Missing inputs, credentials, an empty mining pool, leakage, CSV failures, silent-drop, or AMP allocation mismatch can halt the loop without automatic retry.
How do you install this skill?
Install the specific skill with the command documented in the repository README: npx skills add nvidia/skills --skill tao-run-deft-aoi --yes. You may target a client, for example: npx skills add nvidia/skills --skill tao-run-deft-aoi --agent codex. The skill becomes available when the agent next loads relevant skills.
How do you use this skill?
Prompt the agent with “run the DEFT loop” or “fine-tune my AOI ChangeNet model until FAR is below 0.1% at recall=100%.” You must provide max_iterations. If you provide only a time limit, the workflow converts it into an estimated iteration count and surfaces that estimate in pre-flight. The agent shows the Pre-Flight Summary and waits for explicit approval; after approval, it continues without step-by-step confirmation until the KPI target, iteration limit, or an unrecoverable hard stop is reached.