Module 2 • Lesson 1150 mins

Audit Data Readiness Across 4 Dimensions: Volume, Quality, Label, and Freshness

Overcome the 'we have data' illusion across 4 audit pillars, detect latent historical bias, and avoid the treacherous Proxy Label Trap.

Audit 4 dimensions: Volume, Quality, Label Ground Truth, Freshness
Detect the Proxy Label Trap in historical business logs
Evaluate data drift resulting from operational business shifts

Audit Data Readiness Across 4 Dimensions: Volume, Quality, Label, and Freshness

When evaluating AI feasibility, hearing "We have plenty of data" from engineering or data teams is frequently the precursor to expensive product failures. Having gigabytes of logs in a data warehouse and possessing data that meets production readiness criteria for training or grounding AI are fundamentally different realities.

Running Example: FinTrack Logistics aiming to build an AI model that predicts at-risk delivery delays to trigger proactive rerouting. The data team reports: "We have 5 years of logs covering 20 million shipments."

TIẾP CẬN CŨ / LỖI THỜI
Data Illusion: 'We have 5 years of logs'
Massive volume masking noisy, biased, and stale labels
CHUẨN AI PM / HIỆN ĐẠI
4-Pillar Data Readiness Audit
Rigorous audit across Volume + Quality + Ground-Truth Label + Freshness
DATA FEASIBILITY AUDIT4 Data Criteria & Cold Start Ladder

Audit Data Feasibility: 4 Non-Negotiable Data Criteria

Audit 4 Data Dimensions on FinTrack:
1. VOLUME & DIVERSITY

Only 120 historical fraud cases (severe class imbalance ratio 1:10,000).

2. LABEL QUALITY

No objective ground truth; CS tagged dispute tickets with 30% label inconsistency.

3. FRESHNESS & DRIFT

Adversarial fraud tactics evolve rapidly; models trained on 3-month-old data fail.

4. PRIVACY & COMPLIANCE

Requires exporting raw device fingerprints and biometric traces to third-party endpoints.

Data Feasibility Principle: Models cannot learn what does not exist in data. Audit across the 4 axes (Volume, Label Quality, Freshness, Privacy) and climb the Cold Start ladder rather than stalling for perfect data.