Data Analyst AI Assistant: A Semantically Armored Multi-Agent Framework and Deterministic Execution Engine for Autonomous Data Hygiene and Executive Business Intelligence

Enterprise data preparation remains a severe bottleneck in modern analytics pipelines, consuming up to 80% of engineering effort due to heterogeneous file formats, schema fragmentation, disguised missingness, and domainspecific anomalies. Existing rule-based scripts (e.g., naive Pandas dropna pipelines) suffer from catastrophic row deletion on sparse tables, whereas unconstrained single-pass Large Language Model (LLM) wrappers suffer from context-window saturation on wide schemas, primary-key corruption via blind imputation, and unverified statistical distribution drift. To resolve these limitations, we present the Data Analyst AI Assistant, an end-to-end multi-agent data engineering and business intelligence framework that decouples probabilistic semantic planning from deterministic local execution across four layers: (1) a Content-Sniffing Universal Ingestor and 11 Vectorized Mathematical Sieves that profile 10 file formats while compressing prompt token payloads by up to 97.1% via chaos-score ranking; (2) a 5-Dimensional Quota-Aware Smart Router orchestrating 14 specialized micro-agents with ChromaDB semantic recipe caching; (3) a 55-tool Deterministic Domain-Specific Language (DSL) Execution Engine enforcing high-cardinality Semantic Armor (>95% uniqueness lock), non-destructive Quarantine Vault isolation, and Two-Sample Kolmogorov-Smirnov (KS) Statistical Drift rollbacks (p < 0.05, D_KS > 0.10); and (4) an AST-guarded, self-healing Executive Analytics Agent that deterministically maps cross-file Primary/Foreign Key (PK -> FK) graphs and compiles vector SVG executive dossiers. Evaluated across 28 benchmark datasets (107,085 records) and a 500-trial programmatic fault-injection suite, the proposed architecture elevates mean data quality to 98.88/100 (100.0% accounted retention, p50 local execution latency of 27.8 ms at 1.5K rows and 611.3 ms at 50K rows), outperforming Naive Pandas (86.42/100, 46.85% retention), the Deterministic Shield alone without LLM agents (92.15/100), and Single Monolithic LLM baselines (91.48/100, 75.0% PK integrity).