07/2026 - 08/2026
Aivocat
Local, sovereign AI for a law firm
- Designed and built single-handedly
- Architecture designed by me
Lawyers cannot hand their clients' documents to ChatGPT: professional secrecy. They needed an assistant able to analyse a case file and answer with cited sources, on an ordinary office PC (RTX 3060), with no internet connection at all.
Responsibilities & architecture
- Complete architecture, from the client's need to the installation package, delivered fully autonomously
- RAG architecture: multi-format ingestion with OCR, embeddings (bge-m3), hybrid Qdrant + BM25 search, cross-encoder reranking
- Integration of the Qwen 3.5 9B LLM through Ollama, sharing 12 GB of VRAM between embeddings, reranker and model
- Document search across legal texts (LEGI/DILA open data), queried alongside the case documents
- Legal data confidentiality: 100% offline, case files read-only from a USB drive, signed licence bound to the machine
Results
- About 650 automated tests for about 15,000 lines of Python
- Automated evaluation “jury”: citation accuracy and correct-answer rate measured on test case files
- Answers that quote the exact passages, or abstain when nothing supports them (thresholds calibrated on real measurements)
- Delivered as a fully offline installation package (14 GB), verified with SHA-256 checksums
- Python
- FastAPI
- Qdrant
- Ollama
- Qwen 3.5
- RAG
















