Hi, I'm Mayssa AI Engineer

I build production LLM systems in Monastir, Tunisia: retrieval, agents, and the evaluation that keeps them honest.

See my work Download CV
Portrait of Mayssa Ben Romdhane, smiling, arms crossed
RAG LLM agents Computer vision

2 years building LLM systems that run in production, for 10 HealthTech platforms

Where I've built

  • HealthCare Innovation
  • Continental
  • Monastir University
    Hospital
  • WISI
  • ENISo

Ask about my work

My job is building systems that answer from sources. This page works the same way: pick a question and every claim in the answer links to where it is shown.

What do you build at HealthCare Innovation?

One on-prem stack serving 10 HealthTech and HSE platforms

I design the AI stack behind 10 HealthTech and HSE platforms: the RAG architecture, agent orchestration and the evaluation harnesses 1. The main system is a sovereign RAG that runs on-premises, with open-weight LLMs on dedicated GPUs 2.

  1. Experience · AI Engineer, HealthCare Innovation
  2. Project · Sovereign on-prem RAG

Answered from this page only

About me

I design the AI behind 10 HealthTech platforms: self-hosted LLMs, RAG and agents, with grounding checks that always run.

Before that I spent a year on computer vision R&D at Continental in Germany, and I trained as an engineer in intelligent systems at ENISo. I work in Arabic, French and English.

The Ribat of Monastir seen from its tower, with the town and the sea behind
Monastir, where I live and work
faster answers on the on-prem RAG system
−40%1
documents searchable with hybrid retrieval
10k+1
precision, zero-shot road segmentation
99%2
labeling cost, after auto-annotation
−70%2

Projects that shipped

Four systems, what they changed, and what they run on.

Sovereign on-prem RAG

HealthCare Innovation, Monastir · 2026

Health data can't leave the building, so the whole system runs on-premises: open-weight LLMs on dedicated GPUs, hybrid semantic and keyword search, and a source citation on every answer.

−40%
latency
+25%
relevance
10k+
documents

Qdrant · BM25 · RAGAS · Langfuse · open-weight LLMs

Internal assistantOn-prem · GPU

Which PPE is required for confined-space entry?

A gas detector, a harness with a retrieval line and a supplied-air respirator when oxygen is under 19.5% [1]. A permit must be signed before entry [2].

  • 1 HSE procedure 4.2, p. 7
  • 2 Permit-to-work policy, §3
Faithfulness 0.941.8 s
Interface sketch with sample content

Agents with a human in the loop

HealthCare Innovation, Monastir · 2026

Multi-agent workflows in LangGraph that stop for a person before anything important goes out, recover when a step fails, and are traced from start to finish.

HITL
validation
End-to-end
tracing

LangGraph · MCP · Langfuse

Run #2187 · incident reportWaiting for review
  1. PlannerSplit request into 3 tasks
  2. Retriever12 passages from 4 documents
  3. WriterDraft report, 2 citations
  4. CheckerMissing date, retried once
  5. Human checkSafety officer approves before sending ApproveEdit
Interface sketch with sample content

Zero-shot road segmentation

Continental, Lindau, Germany · 2024

Foundation models find the road for driver-assistance systems without labeled training data, and the same pipeline pre-labels camera images and LiDAR point clouds for the annotation team.

99%
precision
−70%
labeling cost
+8%
mIoU

SAM · SEEM · CLIP · PyTorch · camera-LiDAR fusion

road · 0.99
Illustration: a road mask drawn over public dash-cam footage

Image generation with LoRA

Freelance · 2025

A Stable Diffusion and DALL-E app with LoRA fine-tuning for a client's style, served by FastAPI behind a React interface, with GPU inference tuned for speed.

<5 s
per image

LoRA · Stable Diffusion · FastAPI · React · Docker

StudioLoRA: client-style-v3

a ceramic mug on a linen tablecloth, morning light

4 images4.6 s
Interface sketch · her generated samples go here

Earlier, at university

  • Early melanoma detectionVGG16 and ResNet-50 with transfer learning, 2023
  • ECG classificationSpectrograms and a CNN, 2022
  • Patient-specific dose predictionMonastir University Hospital, validated by its physicians, 2021

How I build a RAG system

The same loop behind every project above: get the right passages, answer only from them, and measure it all the time.

Documents are chunked and embedded, searched with hybrid retrieval, answered by an open-weight LLM with citations, and every answer is scored and fed back DocumentsPDFs, procedures Chunk + embedby section Hybrid searchQdrant + BM25 Open-weight LLMon our GPUs Answerwith citations RAGAS · LLM-as-judge · Langfuse

  1. Retrieve

    Clean the sources, chunk them along their real structure, and search by meaning and by keyword together so exact terms like drug names are never missed.

  2. Reason

    Pick the model and the agent design that fit the task, the budget, and where the data is allowed to live. Often that means open weights on our own hardware.

  3. Evaluate

    Versioned reference sets, RAGAS and an LLM judge score grounding and hallucinations on every release, and Langfuse traces show why an answer went wrong.

Generative AI
LangChain, LangGraph, MCP, LoRA/QLoRA, Hugging Face, Qdrant, ChromaDB
Computer vision
PyTorch, OpenCV, YOLO, U-Net, SAM, CLIP, ONNX, TensorRT, 3D LiDAR
Evaluation
RAGAS, LLM-as-judge, Langfuse, versioned reference datasets
MLOps
Docker, Kubernetes, FastAPI, Redis, AWS, CI/CD
Languages
Python, C++, TypeScript, SQL

Experience

  1. AI Engineer, RAG and agentic systems

    HealthCare Innovation · Monastir

    AI stack for 10 HealthTech and HSE platforms: RAG architecture, agent orchestration, evaluation harnesses.

  2. Generative AI Engineer

    Freelance · Monastir

    Image generation app with LoRA fine-tuning, FastAPI and React, under five seconds per image.

  3. Computer vision R&D, final-year project

    Continental · Lindau, Germany

    Zero-shot road segmentation for driver assistance, and auto-annotation of images and point clouds.

  4. Engineering intern

    WISI · Monastir

    Adaptive multi-bitrate streaming and a local test server with Docker and Node.js.

  5. Bachelor's final-year project

    Carthage Solutions with Monastir University Hospital

    A patient-specific dose prediction model, validated by the hospital's pharmacology department.

Education

Engineering diploma, applied computer science, intelligent systemsNational Engineering School of Sousse (ENISo), 2021–2024

Bachelor's degree in computer scienceFaculty of Sciences of Monastir, 2018–2021

Notes from the work

Short write-ups on what went wrong and what fixed it.

A library card catalogue drawer open on a card that says: ask the librarian

Evaluating RAG without fooling yourself

Why a high retrieval score can still hide wrong answers, and the three checks I now run on every release.

 A punched computer card

Open-weight LLMs on your own GPUs

What it costs, and what it buys you when data can't leave the hospital.

 A country road winding between green hills

Zero-shot segmentation, in practice

Where foundation models were good enough on roads, and where they weren't.

One CV, two languages

One page each, with the same facts as this site.

CV in English CV en français

Questions recruiters ask

Anything else, write to me. I answer within a day.

Contact
What kind of role are you looking for?

AI engineering on RAG, LLM agents and evaluation, and computer vision where it meets them.

Where are you based?

Monastir, Tunisia. Open to remote work and to relocating: to confirm.

Which languages do you work in?

Arabic (native), French (C1) and English (C1).

Can I see code?

Some of it is on GitHub. Employer systems are private; I can walk you through them in a call.

When can you start?

Notice period to confirm.