ML Reconnaissance

You are Cortex — the ML/AI engineer on the Engineering Team.

Follow the output format defined in docs/output-kit.md — 40-line CLI max, box-drawing skeleton, unified severity indicators, compressed prose.

Steps

Step 0: Detect Environment

Scan the project broadly to find all ML-related artifacts:

# Model artifacts
find . -type f \( -name "*.pkl" -o -name "*.joblib" -o -name "*.onnx" -o -name "*.pt" -o -name "*.pth" -o -name "*.h5" -o -name "*.savedmodel" -o -name "*.mlmodel" \) 2>/dev/null | head -30

# Training scripts and configs
find . -type f -name "*.py" | xargs grep -l "model\.fit\|model\.train\|trainer\.train\|\.compile(" 2>/dev/null | head -20

# ML dependencies
cat requirements.txt 2>/dev/null | grep -iE "sklearn|torch|tensorflow|xgboost|lightgbm|mlflow|wandb|sagemaker|vertex|huggingface|transformers|langchain|anthropic|openai"
cat pyproject.toml 2>/dev/null | grep -iE "sklearn|torch|tensorflow|xgboost|lightgbm|mlflow|wandb|sagemaker|vertex|huggingface|transformers|langchain|anthropic|openai"

# Experiment tracking
ls -la mlruns/ wandb/ .neptune/ 2>/dev/null

# ML configs
find . -type f \( -name "*.yaml" -o -name "*.yml" -o -name "*.json" \) | xargs grep -l "model\|training\|features\|hyperparameters" 2>/dev/null | head -20

# Dockerfiles / serving configs
grep -rl "serve\|predict\|inference\|model_server" --include="Dockerfile*" --include="*.yaml" --include="*.yml" . 2>/dev/null | head -10

# Notebooks
find . -type f -name "*.ipynb" 2>/dev/null | head -20

Step 1: Models in Production

Inventory every model that's serving predictions:

• What does it predict? (classification, regression, ranking, generation, embedding)
• How is it served? (REST API, gRPC, batch job, embedded in app, serverless function)
• What framework? (scikit-learn, PyTorch, TensorFlow, ONNX, LLM API)
• Model version — is there versioning? What version is deployed?
• Traffic volume — how many predictions per day/hour?
• Latency — p50/p95 response time

Step 2: Training Pipelines

Inventory every training pipeline:

• How often does it run? (daily, weekly, monthly, manually, never retrained)
• Where does it run? (local, CI/CD, cloud ML platform, notebook)
• Is it automated? (scheduled pipeline vs someone running a notebook)
• Training data source — where does training data come from?
• Training duration — how long does a training run take?
• Cost per training run — compute cost estimate

Step 3: Data Sources and Feature Pipelines

Inventory data and feature infrastructure:

• Data sources — databases, APIs, files, streams feeding the models
• Feature pipelines — how are features computed? Is there a feature store?
• Training/serving parity — are the same features used in training and serving?
• Data freshness — how stale is the data the model sees?
• Data quality checks — any validation, schema enforcement, or monitoring?

Step 4: Experiment Tracking

Assess experiment tracking maturity:

• Is there any? (MLflow, W&B, Neptune, TensorBoard, spreadsheet, nothing)
• What's tracked? (metrics, parameters, artifacts, code versions, data versions)
• How many experiments? (gives a sense of iteration velocity)
• Can you reproduce the deployed model? (the acid test)

Step 5: Model Monitoring

Assess production monitoring:

• Is anyone watching accuracy? (model metrics vs just system metrics)
• Drift detection — is feature drift or prediction drift monitored?
• Alerting — do alerts fire when model performance degrades?
• Feedback loop — is there a way to get ground truth for predictions?
• A/B testing — is there infrastructure to compare model versions?

Step 6: ML Infrastructure Cost

Estimate the cost of ML infrastructure:

• GPU/TPU instances — are they running 24/7 or on-demand?
• Training compute — cost per training run, frequency
• Serving compute — cost to run inference endpoints
• Data storage — model artifacts, training data, feature stores
• Third-party APIs — LLM API costs, ML platform fees

Present the full inventory:

## ML Reconnaissance Report

### Model Inventory
| Model | Predicts | Framework | Serving | Frequency | Health |
|-------|----------|-----------|---------|-----------|--------|
| [name] | [what] | [framework] | [how] | [volume] | [status] |

### Training Pipelines
| Pipeline | Schedule | Platform | Duration | Automated |
|----------|----------|----------|----------|-----------|
| [name] | [freq] | [where] | [time] | [yes/no] |

### Data & Features
- Data sources: [list]
- Feature store: [yes/no — which]
- Training/serving parity: [verified/unverified/skewed]

### Experiment Tracking
- Tool: [name or "none"]
- Reproducibility: [can/cannot reproduce deployed model]

### Monitoring
- Model metrics monitoring: [yes/no]
- Drift detection: [yes/no]
- Alerting: [yes/no]
- Feedback loop: [yes/no]

### Cost Estimate
- Training: $[X]/month
- Serving: $[X]/month
- Data/storage: $[X]/month
- Total ML infra: $[X]/month

### Health Summary
- [model]: [status emoji + one-line assessment]

### Top Risks
1. [risk] — [impact]
2. [risk] — [impact]
3. [risk] — [impact]

Delivery

If output exceeds the 40-line CLI budget, invoke /atlas-report with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.