Methodology
Architecture and operating constraints for a personal portfolio project focused on AI engineering and data engineering.
Data source strategy
Synthetic mode remains the deterministic default. Official mode uses a separately ingested, checksummed 250-record DGA/data.go.th snapshot with record-level attribution; the two modes are never aggregated.
AI design
The current public implementation uses deterministic, evidence-bound generation and cached summaries. External GenAI remains an optional extension point rather than a runtime requirement.
Cost control
The demo works without private API keys. Embeddings use a local deterministic fallback, summary generation is cached, and deployment targets free-tier Vercel, hosted FastAPI, and Supabase PostgreSQL.
Privacy boundary
The bounded official snapshot excludes supplier names and legal identifiers. Public data is not proof of wrongdoing, and this project does not rank agencies or vendors as suspicious.
Bounded evidence
The snapshot is used to demonstrate acquisition, mapping, quality checks, provenance, retrieval, and citations. It is incomplete, may become stale, and does not represent the entire Thai procurement system.