10/22/2026 updated
AR
100 % available
AI consultant and engineer (LLM, prompt engineering and evaluation, RAG, AI agents, document AI)
Dubai, United Arab Emirates
Only remote
About me
AI consultant and engineer based in Dubai, founder of ARELION FZCO (AI consultancy). Available now, remote only. LinkedIn: https://www.linkedin.com/in/antoninribeaud/
AI AgentsRAGLLMOpenAIGeminiVertex AILangChainArtificial IntelligenceOptical Character Recognition (OCR)Information RetrievalPython (Programming Language)Machine LearningNatural Language ProcessingTypeScriptGoogle Cloud
I hand the repetitive part of a team's work to AI, with a person approving each output, so people get those hours back for the work that needs judgment. I build the systems myself: 12 years of shipping software, today from Dubai. In production: search across 100M+ pages of R&D documents for a global enterprise, a newsroom assistant over 500,000+ articles, legal research with checked citations for a regulated client. I also act as fractional CTO and audit teams, code security and vendor contracts.
Prompt engineering and LLM output quality (public case studies on arelion.dev, clients anonymized):
- Prompt engineering: I write and tune the prompts of production LLM features (answers with citations, structured extraction, agents with tools).
- Evaluation: every prompt or model change is re-scored on cases mined from production traffic, with an LLM judge, before users see it (Google ADK EvalSet, Vertex AI Gen AI Evaluation). arelion.dev/case-studies/agent-eval
- Structured output: model outputs must pass strict Pydantic schemas, and generated SQL is parsed (sqlglot) and limited to the user's rights. Output that fails is rejected. arelion.dev/case-studies/output-contracts-in-production
- Grounded answers: code checks every citation against the source corpus and removes what it cannot verify. arelion.dev/case-studies/legal-research-assistant
- Prompt injection: untrusted text cannot reach a privileged action or an outbound channel without passing deterministic code. arelion.dev/case-studies/prompt-injection-defense
- Model comparison: open benchmark of 14 OCR engines, LLMs included (Gemini, GPT, Claude, Qwen, Mistral), on Arabic documents. arelion.dev/arabic-ocr
Skills: LLM and generative AI, prompt engineering, LLM evaluation (golden sets, LLM as a judge, regression tests), structured output (JSON, Pydantic), RAG and enterprise document search, AI agents (Google ADK, LangChain, MCP), document AI and OCR (including Arabic OCR), Python (FastAPI), TypeScript, Google Cloud Platform (Vertex AI, Cloud Run, BigQuery), Pinecone, LiteLLM. Models: Gemini, Claude, GPT, Mistral, Qwen.
Remote only, based in Dubai (UTC+4). Languages: English, French.
Experience
Professional experience10+ years
Freelancer since2014
Languages
EnglishFluentFrenchNative speaker
Project history
My freelance work since January 2026, through my company in Dubai.
- Legal research assistant for a regulated client: code checks every citation against a versioned legal corpus, and any reference it cannot verify is removed before the reader sees it. Case study: arelion.dev/case-studies/legal-research-assistant
- Newsroom assistant for a national news publisher (under NDA): staff ask questions and get answers from 500,000+ archived articles. Case study: arelion.dev/case-studies/newsroom-second-brain
- Open benchmark of 14 OCR engines, LLMs included (Gemini, GPT, Claude, Qwen, Mistral), on Arabic documents. Code and scores are public: arelion.dev/arabic-ocr
- Fractional CTO for the same publisher: platform rebuilt, and the full article archive migrated with no downtime for readers.
Stack: Python, Google ADK, Gemini, Vertex AI Search, PostgreSQL, MCP.
- Built the document AI platform used by L'Oréal R&D teams.
- Designed the pipeline that fetches, OCRs, vectorizes and indexes 100M+ multilingual pages.
- Delivered semantic search and question answering on top of the index.
- Built the microservices on Google Cloud, with the infrastructure in Terraform.
Stack: Python, LLMs, RAG, Google Cloud, Terraform.
Document AI product I built alone, from architecture to go-to-market. IP and code sold in 2025.
- Teams ask a question to their own files and get the answer with its source, scanned PDFs and several languages included.
- OCR, parsing and categorization of documents, then answers through RAG, with structured JSON output and sources.
- Prompt engineering, and vector search with query rewriting and expansion.
- Works with OpenAI, Claude or Mistral models, on a single server.
Stack: LLM APIs (OpenAI, Claude, Mistral), RAG, vector search, OCR.