Zakaria Si Salah

Since October 2025, I have been working as a self-employed professional on two parallel projects.

Project 1 – AI Agentic R&D & LLMOps

I design and deliver operational agentic systems for companies looking to automate their business processes with generative AI. Specifically: agents capable of multi-step reasoning, tool usage (web search, databases, business APIs), complex task planning, and self-correction without human intervention. I master the most advanced current patterns: multi-agent architectures with role specialization, agentic workflow orchestration, human-in-the-loop (HITL) supervision, drift prevention guardrails, and hallucination management. In parallel, I build high-precision enterprise RAG pipelines: semantic segmentation, hybrid reranking, automated response quality evaluation. I also deploy open-source LLMs on private infrastructure with fine-tuning on proprietary data and full observability. What I deliver: a documented, testable, maintainable system, not a prototype.

Project 2 – Restaurant SaaS

Following a specific request from a restaurant client looking for an alternative to existing solutions, I designed and developed a solution from scratch: an all-in-one SaaS ecosystem for commercial restaurants. The solution integrates a touchscreen kiosk software, a kitchen display system (KDS), an online click & collect ordering system, a loyalty program, and multi-establishment management. Three key differentiators: 0% commission, 100% guaranteed offline operation (no downtime in case of network outage), and native tax compliance (transaction cryptographic chaining, automated Z reports, FEC export). Customizable to the restaurant's branding: kiosks, receipts, ordering website: the end customer only sees the restaurant's branding; the product is entirely in the establishment's image. Designed to scale.

Led the design of an enterprise RAG system for SNCF development teams, reducing information retrieval time by 70% and automating traceability between specifications, code, and documentation.

Designed a 3-level architecture: an AI engine orchestrating agents (LangGraph state machines), a responsive Next.js interface, and hybrid storage (Weaviate for vector data + Neo4j for graph + local SQLite) for cross-repository semantic search with preserved context.

Implemented a FastAPI backend with 6 AI agents: code explanation (syntax/semantic analysis), cross-microservice search (dependency graph), pattern generation, technical documentation Q&A, educational tutorials, and bidirectional code-documentation traceability.

Developed an MLOps ingestion pipeline: sophisticated PDF extractors (tables, images, hierarchy), code extractors (dependencies, component graph), intelligent context segmentation, and optimized chunking for vectorization.

Optimized vectorization for the technical domain using specialized embeddings + high-performance Weaviate indexing + Neo4j component relationships, ensuring automatic consistency between the codebase and documentation on each commit.

Optimized LLM usage: intelligent routing to the appropriate agent, secure local storage with zero external leakage, indexing tailored for technical content, reducing tokens by 40%.

Developed a native Tauri application (Rust/TypeScript) with a cross-platform installer (Win/Linux/Mac), offering native performance and local security without cloud dependencies.

Green IT: Used a local DeepSeek model for development/testing, optimized LLM production (OpenAI/Anthropic) for contextualization tokens, reducing energy footprint by 60%.

Managed 4GB GPU resources: quantized models, cached embeddings, automatic memory release, and dynamic load adaptation.

Analyzed difficulties in onboarding new hires (contribution time), knowledge loss of old code, and post-delivery bugs lasting 3+ years.

Implemented GitLab CI/CD pipelines for 16 microservices in production, integrating compilation, automated testing, packaging, and end-to-end deployment, reducing deployment time from 30 minutes to 5 minutes.

Containerized and orchestrated services using Docker and Kubernetes (Minikube for local, Rancher for production), configuring 1 cluster for multi-environment orchestration.

Automated deployment processes using structured Helm Charts, eliminating 90% of manual interventions and significantly limiting human errors.

Contributed to setting up a distributed infrastructure on embedded box hardware using Harbor (private Docker registry) and Rancher for centralized cluster and image management, within a context of strong hardware constraints.

Designed an Nginx load balancing mechanism to ensure front-end/back-end communication under embedded OS constraints, improving Kubernetes platform availability and resilience.

Developed Python automation tools for maintenance and critical package updates, reducing typical processing time from about 1 hour to 10 minutes, freeing up team capacity.

Participated in the architectural redesign of an embedded component into a lightweight version, reducing resource consumption (CPU, memory, storage).

Contributed to an embedded microservices project for train systems in Golang and Vue.js, developing high-performance backend services and robust operator interfaces adapted to field use.

Implemented an advanced train localization algorithm capable of maintaining coherent position during GPS signal loss and resynchronization, improving overall position tracking reliability.

Designed and maintained comprehensive unit and integration test suites covering key business and technical scenarios, increasing code coverage from 60% to 85% and enabling early detection of regressions.

Participated in critical production incident management (in-depth diagnostics, fix development, deployment), contributing to an estimated 25% improvement in system stability.