Badr Eddine

I owned an end-to-end, learning-to-rank experiment, handling the research, design, implementation, testing, and deployment phases. Deployed the ranking model across media categories for several marketplaces. The A/B testing results indicated an increase of more than $30 million in Gross Merchandise Value (GMS).

A Luxembourgish language-learning app where a RAG pipeline grounds every explanation in a vetted corpus and constrains output to the learner's level (A1/A2 CEFR), so lessons never drift above their level or invent vocabulary. I added a faithfulness self-review layer that checks each response against the retrieved context before it reaches the learner, catching hallucinations and off-level content. Evaluation was first-class throughout: groundedness and faithfulness scoring, context precision and recall, and LLM-as-judge to catch regressions. The result is a tutor whose answers stay grounded, on level, and safe to ship, plus a reusable retrieve-generate-verify pattern for any LLM product that needs to be reliable.

An AI platform for job discovery and interview prep that I co-founded and built end-to-end. The matching engine parses a candidate's resume (OCR), embeds and semantically matches it to roles, then applies an LLM-based fit-scorer and re-ranker conditioned on the CV. In other words, a recommendation engine for people and opportunities, with personalization driven by each profile. It also runs a real-time AI mock interview: live voice (OpenAI Realtime), on-camera body-language feedback (Gemini Vision), and instant scoring. We grew to 95 paying subscribers on a two-person team, and I own the full stack: Next.js, FastAPI, OpenAI, Gemini, and GCP Cloud Run.