Abed Matini

Abed Matini is an AI engineer and backend developer building conversational systems that remain affordable, inspectable, and useful when handling real customers or sensitive documents. His work ranges from KAIIA, KFC South Africa’s interactive assistant, to local-first document AI that keeps organizational knowledge under tighter control.

Matini began in web development, including WordPress and Joomla projects, before moving into backend engineering. He earned a Master of Technology in Information Technology at Cape Peninsula University of Technology, researching how stress, chatbot interactions, and emotional responses influence online purchasing. In 2024, he co-authored research on chatbot interactions and purchase intention.

At Ogilvy South Africa, he contributed as a developer to KFC’s award-recognized “What should we fry?” campaign and handled backend development for KAIIA, which extended an advertising character into interactive cinema, drive-through, and online experiences. His later work concentrates on retrieval-augmented generation, privacy, and the operational choices behind dependable document assistants.

His public build-local-rag-in-30-days project develops a local document assistant using Ollama, Docling, Chainlit, and PostgreSQL’s pgvector extension without requiring a GPU. Other owned repositories explore chatbot development, retrieval datasets, image captioning, and Drupal environments.

How he builds dependable document assistants

• Structure documents before retrieval. Matini converts PDFs, office files, and images into inspectable Markdown, then chooses heading-based, paragraph-based, fixed-window, or sentence-based chunks according to the material. For structured product documentation, he favors heading-based chunks because citations remain intelligible.

• Combine semantic understanding with exact matches. His hybrid retrieval with Reciprocal Rank Fusion merges vector and keyword search inside PostgreSQL. This preserves conceptual relevance while recovering details that similarity search can miss, including product identifiers and medication names.

• Reserve model calls for actual reasoning. Dates, calculations, rule checks, and straightforward lookups belong in testable Python functions. Fixed retrieval pipelines offer more predictable latency and stronger control for compliance-sensitive applications than unnecessary agent loops.

• Make failures visible and block unsafe requests early. Matini pairs retrieval observability with user-session traces, source citations, consent controls, and prompt-injection safeguards applied before requests reach a model. His AI Engineer World’s Fair demonstration combined FastAPI, React, PostgreSQL, Docker, Ollama, and Langfuse in an employee-handbook assistant built around these principles.