Yassine Derbel — AI Engineer

I make AI useful, not just impressive.

AI engineer with a background in banking automation, QA tooling and medical vision. I've built an on-prem RAG assistant that cut test-failure analysis to under 2 minutes, automated core-banking SWIFT flows, and shown that small language models can match LLMs on real enterprise tasks for a fraction of the compute.

RAG & document AI

chunking · embeddings · eval

LLM agents

tool use · citations · tracing

Small / on-prem models

SLMs · Ollama · fine-tuning

Computer vision

YOLOv8 · segmentation · real-time

Automation & backends

FastAPI · Postgres · CI/CD

01QA · EnterpriseOn-prem

Qualitest

A RAG assistant that reads a failed automated test and tells engineers why it broke and how to fix it.

Three-phase dataset: human labelling → RAG enhancement → AI augmentation over Playwright & Puppeteer failure reports.
Test codebase chunked with AST parsing, embedded with nomic-embed and indexed in ChromaDB.
Llama 3.2 via Ollama, fully on-premise; FastAPI + Postgres jobs, live dashboard, GitLab & Jenkins integration.

LangChainOllamaChromaDBFastAPIJenkins

<2 minper failure analysis

01failure report ingested

02AST chunks retrieved · k=8

03similar past fixes matched

04root cause + fix steps

05engineer feedback → regenerate

02Construction · AECAgent

Mensura

Chat with DWG/DXF construction drawings — counts, lengths, areas, quantity takeoffs and BoQ reconciliation.

The LLM never does arithmetic: it calls 20+ deterministic geometry tools built on ezdxf, Shapely and Postgres.
Every number is cited back to entity handles you can click to highlight on the drawing.
8-layer architecture: ingest → normalise → geometry → tools → agent → BoQ → reconcile → UI, traced with Langfuse.

Claude tool useezdxfShapelypgvectorReact

20+deterministic tools

user › how many columns on level 2?

tool › list_layers(filter="COL*")

↳ S-COLS-L2 · 148 entities

tool › count_blocks(layer="S-COLS-L2")

↳ 42 · handles [3A1F, 3A20, …]

tool › sum_lengths(layer="A-WALL")

↳ 1,284.6 m

agent › 42 columns, 1,284.6 m of walls [cited]

03Generative videoOpen source

Sinister Synthesis

A fully automated studio that turns a story into a published YouTube Short — with zero paid APIs.

Script by Llama 3.1 (Ollama), voice by edge-tts / Piper, visuals by SDXL.
Ken Burns motion, word-level Whisper captions, FFmpeg compositing.
Uploads itself via the YouTube Data API — 8 stages, no human in the loop.

Llama 3.1SDXLWhisperFFmpegPython

$0API spend per video

①source

②script

③voice

④visuals

⑤motion

⑥captions

⑦edit

⑧upload

04Assistive techReal-time CV

CareGenius

A webcam coach that guides autistic children through putting on their shoes — and shows caregivers their progress.

Two YOLOv8-seg models (foot, shoe); coverage from pixel-level mask overlap, not bounding boxes.
Gemini assesses the shoe's state and jumps to the right step; Deepgram voice gives child-friendly prompts and praise.
Caregiver dashboard with session analytics and SMS alerts when help is needed.

YOLOv8-segGeminiDeepgramSupabaseNode

2 fpslive mask analysis

0%foot in shoe

shoe offpartial ≥15%fully in ≥60%

05Robotics · NLP🥇 1st in Africa

Beemo

A companion robot that talks with autistic children and reads their emotions in real time.

Conversation model built on LLaMA 3 70B + LangChain, tuned on real therapist–child sessions with Tunisian autism associations.
Facial emotion recognition at 92% accuracy to adapt the conversation.
1st place, Africa to Silicon Valley Hackathon · ESPRIT Best AI Project.

LLaMA 3 70BLangChainCNNRobotics

#1Africa · ASV Hackathon

happycalmsurprisedsad

Clients keep coming back.

“Yassine is an amazing developer and an even better person. The project was complex…”

★ 5.0 · AI image rotation & cropping tool

“He communicated clearly, made changes quickly based on feedback, and showed real dedication to delivering the project well.”

★ 5.0 · Full-stack AI developer for a GenAI MVP

Banks, platforms & startups.

Nov 2026 — now

Freelance AI Engineer

Independent · Upwork & direct

Designing and shipping RAG, agentic and computer-vision systems for clients in construction, health, e-commerce and media.

Jan — Oct 2026

Data Science Engineer

VALUE Digital Services · on mission at BIAT

Built an automation system for Temenos T24 core-banking processes, including a robot that processes SWIFT MT103 payment messages, deployed on the bank's internal vSphere.
Automated printing and mailing workflows end to end, from Outlook ingestion to batch printing.
Led an SLM research project showing small language models can compete with LLMs on targeted enterprise tasks at a fraction of the compute — fully on-premise.

Feb — Sep 2025

AI Engineer · end-of-studies

NeoXam

Built Qualitest, an on-prem RAG assistant that diagnoses failed automated tests and recommends fixes — under 2 minutes per report.
AST-based code chunking, ChromaDB, Llama 3.2 via Ollama; FastAPI/Postgres backend with GitLab & Jenkins CI integration.

Jun — Sep 2024

Data Scientist · intern

Tabaani

Built the YouTube scraping and speech-to-text pipeline and Llama 3 activity extraction behind the Middle East's first automated travel-planning chatbot.

Jun — Aug 2023

Computer Vision Engineer · intern

InTech Solutions

OpenCV/CNN pipeline turning ECG scans into tabular data — myocarditis diagnosis from days to seconds — plus a heart-failure prediction model.

2020 — 2025

Engineering Degree, AI

ESPRIT, Tunis

Final project: A RAG Approach for Quality Assurance Optimization. Chairman of DeepFlow – AI For All (2023–24).

Yassine Derbel — Freelance AI Engineer