Benoît Prioux

@binout.bsky.social

Enthusiast Software Developer 💚 @avec_alan 📣 Alan Recrute https://jobs.ashbyhq.com/alan?departmentId=8ba0fb44-ec79-4731-bf45-d65a821e24f0

Few-shot examples are only as good as their labels. Ours were hallucinating values from memory. The LLM learned to do the same. Part 2 of the doc processing series: curated datasets, backtesting every change and separating parsing from human knowledge medium.com/alan/lessons...

Lessons from Running an LLM Document Processing Pipeline in Production

We evolved from text-only to multimodal, built curated reference datasets for RAG and developed an evaluation framework

medium.com

Ever wondered how Alan reimburses health claims so fast? Spoiler: we automated 70% of document processing. We just dropped the first article of our document processing series — stay tuned for more on how we're scaling this across other countries and use cases. buff.ly/2Lu7aWD

How we reached 70% document processing automation at Alan

We replaced our classic ML pipeline with a unified LLM-based approach using Markdown Transcription and few-shot RAG extraction

medium.com

Loïc Kervran et moi avons l'immense honneur de pouvoir parler à @devoxx.fr cette année 🥳 Nous vous présenterons "All-In sur la plateforme : nouveau Poker @Betclic" ♦️♠️♥️♣️ Un retour d'XP sur les challenges derrière la création (en un an 🤯 ) d'une Poker room à la croissance toujours plus forte 💪

Devs who write about their work/company; talk at conferences - lots of companies in the past used to not care much about this and not value it the least. Feels like the tide is changing: people who do this but are not valued get hunted down by companies who understand this value

"Claude did it almost on its own while I was boiling water for my tea" That's how our engineers now fix flaky tests at Alan. Read more about our journey from manual debugging to AI-powered investigation in our latest article 👇 medium.com/alan/from-gu...

From Gut Feeling to AI assisted: How We Automated Flaky Test Investigation

Over the last year, we’ve improved our flaky tests investigation process, moving from manual debugging to an AI first solution. This…

medium.com

Je crois qu'une des choses les plus utiles que j'ai acquises au fil des années dans ma carrière de développeur c'est de faire une distinction nette entre une préférence et une conviction. C'est absolument indispensable pour travailler en équipe et trouver des compromis.

There are many LLM benchmarks such as MMLU and GSM8k, but they're useless for AI agents. Real agents need to handle database state, tool calling, and multi-turn conversations. Stateful benchmarks show the path forward. New post on agent evaluation 👇

Benchmarking AI Agents: The Challenge of Real-World Evaluation

AI agents need stateful benchmarks. Unlike LLMs, agents interact with databases and users. We explore why and how to evaluate them…

medium.com