Founder and AI engineer - French administrative law in 8 languages
AsterixLab Production RAG Architecture
A production RAG case study covering multilingual legal retrieval, 203,427 vectorized records, two-pass chunking, evaluation, and five resolved failure modes.
Problem, role, and approach
- Problem
- Asylum seekers and immigrants cannot read French administrative law, and low-resource languages like Tigrinya and Pashto break standard retrieval pipelines.
- Role
- Founder and sole builder. Owned architecture, retrieval logic, data pipeline, and CI/CD-gated deployment.
- Approach
- Vectorized 203,427 legal records across four corpora with Mistral-embed and Supabase pgvector, then designed a two-pass chunking strategy to close the tokenization gap for low-resource languages.
Decisions
- Used Mistral-embed with Supabase pgvector for the retrieval layer.
- Introduced a two-pass chunking strategy for low-resource-language retrieval.
- Gated deployment through the tested CI/CD pipeline.
Architecture
- Four legal corpora containing 203,427 vectorized records.
- Sentence-level retrieval followed by paragraph-context reranking.
- Eight-language answer experience backed by CI/CD quality gates.
Evaluation
- 29 of 29 tests green in CI/CD at the reported release state.
- Retrieval parity checked across high- and low-resource languages.
Failure modes
- Standard retrieval pipelines underperformed for low-resource languages.
- Five distinct production retrieval failure modes were identified and resolved.
Tradeoffs
- Two retrieval passes add pipeline complexity in exchange for paragraph context and better low-resource-language retrieval.
Evidence
- 203,427 records vectorized across four legal corpora.
- 29 of 29 tests green in the reported CI/CD state.
- Live product available at chat.asterixlab.fr.
Employer value
Evidence of end-to-end AI platform ownership across product, architecture, retrieval evaluation, debugging, and production delivery.