AI Engineer — RAG · Fine-tuning · LLMOps
Private, production GenAI systems, built end to end and measured before they ship.
Final-year Computer Engineering student in Porto, building private, on-prem AI systems from the backend and queues to the RAG pipeline, the fine-tuned model and the desktop UI. I don't ship an AI change without a number behind it: every retrieval or model change goes through rag-eval-harness before it lands.
At Fashable I built the backend and Azure infrastructure behind multimodal product search, demoed the MVP at NRF 2024 (New York) and saw it ship on the Microsoft Marketplace as a Teams Copilot app.
| Project | What it is | What it shows |
|---|---|---|
| rag-eval-harness | Regression gates for RAG pipelines: hit rate, MRR, keyword recall, refusal and citation checks, p95 latency, threshold gates and per-question diffs between runs. No LLM-as-judge. | Evaluation discipline. The same loop LocalVault runs internally. |
| qlora-8gb-pipeline | JSONL dataset → 4-bit QLoRA adapter (Unsloth + TRL, Qwen 2.5 3B) → GGUF for Ollama, on a single RTX 4060. Digest-pinned Docker image and a provenance manifest per run. | Reproducible fine-tuning under a hard memory budget. |
| swiss-legal-rag · alpha | Local, trilingual RAG over Swiss federal law (Fedlex, DE/FR/IT): hybrid pgvector + full-text search with Reciprocal Rank Fusion, cross-encoder reranking, article-level citations, 33-question gold set. Offline Tauri desktop app. | Multilingual retrieval with citations you can check. |
| YOLOv8 | Retail shelf monitoring on SKU-110K and Shelves: 0.900 mAP@0.5 with Nano at ~3.2 ms per image on an RTX 4060. | More labelled data beat a 3× larger backbone. |
| LocalVault · closed-source | On-prem AI platform: local RAG (pgvector, cross-encoder reranking, semantic cache), QLoRA fine-tuning on an 8 GB GPU, MLflow evals and Langfuse tracing. NestJS + BullMQ backend, Tauri/React desktop app. | The system the three tools above were built for. Case study → |
More case studies, including maiai and nexus-doc-ai, on samuel-carvalho.com/projects.
|
AI / LLM LangChain • pgvector • Ollama QLoRA / Unsloth • Langfuse • MLflow |
Backend & Data NestJS • Prisma • BullMQ PostgreSQL • Redis |
Frontend & Infra React • Tauri • Tailwind Docker • MinIO • Azure |
Also comfortable with Java (Spring Boot), Go and MongoDB.