Mounaim Belkhoumali

I designed and delivered 2 B2B SaaS platforms integrating production LLMs end-to-end, plus a part-time backend engagement on the MapBrain platform. I ensured the transition from AI prototypes (LLMs, autonomous agents) to production-ready applications.

A SaaS platform delivered as a white-label handover to an IP consulting firm. Features a conversational patent-drafting assistant (a guided 4-step chat generating an INPI/EPO draft), a multi-source prior-art search engine (USPTO / EPO / Google Patents), an INPI official-notification response generator, a GDPR-compliant self-hosted voice assistant, and a full multi-tenant back-office, enabling IP consultants to draft and manage portfolios faster.

A document management SaaS for Swiss fiduciaries. A multimodal LLM pipeline extracts and classifies documents (invoices, statements, receipts), cutting manual data entry by 70%, while an email RAG lets accountants query client history in natural language, with Gmail/Outlook integration and Bexio / Crésus / WinBiz connectors aligned to Swiss tax formats, saving hours of manual bookkeeping per client.

An educational SaaS that transforms any document (PDF, YouTube video, audio, slides) into an interactive learning path: summaries, mindmaps, quizzes, flashcards, generated podcasts, and a real-time conversational voice tutor, letting learners absorb material in the format that suits them best. Deployed end-to-end on Azure.

Problem:US clients needed to move from idea to reliable AI product: monitor their online reputation, process complex medical documents, produce editorial content at scale, all without an internal AI team.

Real-time e-reputation monitoring.Continuous multi-source collection (web, social media, forums), sentiment analysis by LLM, action recommendation engine, and automatic email reports. The client moves from impossible manual monitoring to actionable alerts.

Medical AI Assistant. The challenge:long and technical clinical documents where an interpretation error is costly. Architecture combining autonomous agents and sourced RAG: each response is traceable to the exact passage in the original document, an essential condition for practitioners' trust. The tool concretely accelerates clinical decision-making.

Automated editorial chain. End-to-end pipeline:detection of trending topics, generation of articles and visuals (DALL·E), scheduled publication on LinkedIn, X, and WordPress via their APIs. Zero human intervention between topic detection and going live.

Problem:Sales teams manually sifted through multi-page PDF tenders to extract response criteria. A slow, repetitive, and error-prone task, directly impacting the bid response rate.

Technical Solution:RAG system developed with LangChain automatically extracting criteria from PDFs. Since the raw accuracy of a naive RAG was insufficient, I optimized it with advanced techniques: HyDE (generating hypotheses to improve vector search relevance) and RAG Fusion (merging multiple queries). Most importantly, I instrumented quality with RAGAS metrics (faithfulness, answer relevancy, context precision) and conducted systematic benchmarking (chunking strategies, top-k values, embedding models) to select the optimal configuration on real data. Because an AI system without quality measurement is not a production system. Result: 50% less manual work.