SQL path
Counts, rankings, known entities, and structured comparisons use PostgreSQL. The answer is reproducible and its source is explicit.
Case study / Civic data + AI
MONÉLU
A public interface to French parliamentary activity, built from static archives and designed to show where every answer comes from.
646Deputies tracked
5,404Votes analyzed
797K+Vote positions
3,741Indexed chunks
01 / Constraint
The API did not exist.
The Assemblée Nationale publishes useful open data as static ZIP archives rather than a stable REST API. MonÉlu turns those inconsistent exports into a repeatable product: fetch, parse, validate, upsert, model, serve, and refresh without losing provenance.
02 / Data engine
From archive to trusted answer.
The pipeline separates acquisition, transformation, serving, and interface concerns. Retries and idempotent upserts protect ingestion; dbt tests the analytical layer; FastAPI remains stateless; scheduled GitHub Actions keep production current.
AssembléeStatic ZIP exports
IngestionPython parsing + retries
PostgreSQLSupabase + pgvector
dbt + APITested marts + FastAPI
ProductVercel web interface
03 / AI engine
Route first. Generate second.
MonÉlu does not ask a language model to improvise over every question. Deterministic questions route to SQL. Semantic questions retrieve relevant evidence through pgvector before a French-language response is generated.
RAG path
Open-ended questions retrieve indexed evidence, expose the supporting passages when available, and identify RAG as the source.
04 / Engineering view
Governance is part of the interface.
Confidence does not come from polished prose. It comes from observable retrieval, tested models, explicit routing, and a product that distinguishes database facts from generated synthesis. A keyword evaluation score above 0.80 provides a measurable baseline for retrieval quality while the architecture leaves room for better evaluations over time.
05 / Ownership
Built across the whole path.
I designed and shipped the ingestion, dbt models, FastAPI layer, SQL/RAG routing, evaluation baseline, and web product. The goal was not a demo chatbot; it was a public system whose data path can be inspected.
My role
Product architecture, data engineering, AI retrieval, API design, deployment, and interface delivery.
Trade-off
Scheduled batch fits the Assembly's ZIP publishing model. Supabase and Railway keep operations lean while preserving clear service boundaries.
Outcome
A live bilingual product with tested marts, explicit answer routing, daily refreshes, and measurable retrieval quality.