Skip to content

Houssam El Ayadi · AI & Data Systems Engineer

Complex data.Clearer decisions.

I design machine learning models and the data systems that make them useful: data preparation, experimentation, evaluation, and integration into applications.

Explore my projectsDiscuss a project
  • Data engineering
  • RAG & LLMs
  • Automation
  • Full-stack
Scroll to explore

01Capabilities

Four disciplines, one system.

These usually sit in separate teams. My work is connecting them, so data, models and product move together.

  1. 01

    AI systems

    Retrieval-augmented generation and LLM features that are grounded in real data and wired into a product.

    • RAG pipelines
    • Embeddings & vector search
    • LLM integration
    • AI feedback & scoring

    Seen inDocument question-answering pipelineAI content generation platformClinical training simulator

  2. 02

    Data platforms

    Pipelines that take high-volume, messy data and make it searchable, analysable and actionable.

    • Authentication log ingest & parsing
    • Search & aggregation
    • Distributed analysis
    • Dashboards

    Seen inAuthentication log analytics

  3. 03

    Automation

    Multi-step processes with explicit control flow, so behaviour stays predictable when AI is involved.

    • Backend orchestration
    • State machines
    • Evaluation workflows
    • KPI alerting

    Seen inClinical training simulatorAuthentication log analytics

  4. 04

    Full-stack products

    APIs, persistence and interfaces that turn a pipeline into something people can actually use.

    • FastAPI services
    • Data modelling
    • Application UIs
    • Supabase · MongoDB · PostgreSQL

    Seen inAI content generation platformClinical training simulator

02Services

Select a challenge. See the system take shape.

Five kinds of work I take on, from a first AI feature to modernizing a product that already exists. Each one ends with something your team can run.

  • fig. 01inputs → model → review → output

    Concept illustration: AI Applications & Workflows

    When your team loses hours creating, sorting or searching for information, I build AI features and connected workflows that do the heavy lifting, with a person still in control of what goes out.

    Discuss an AI workflow

    What you get

    1. 1.AI-powered interfaces for generating or processing content
    2. 2.Integrations between language models, your data and your existing tools
    3. 3.Human review steps before anything is published or sent

    Typical tools

    • LLMs
    • RAG
    • FastAPI
    • Python

03Selected work

Systems I've designed and built.

An internal product in development, a completed end-of-studies pipeline, a clinical MVP underway, and a document RAG proof of concept. Each case states the problem, what I built, and how far the work went.

01 — FeaturedGenAI · Content generation

Internal product / in development

AI content generation platform

content-platform / new brief
Illustration

Brief

Announce a new feature to existing customers. Friendly, concise, one clear call to action.

Context retrieved

  • brand-voice
  • product-notes
  • past-campaigns

Draft

Generating

[1] brand-voice[2] product-notes
  1. FastAPI
  2. RAG / LLM
  3. Store
  4. AI services
  5. Output
Illustration of a brief being turned into generated content using retrieved context.
Problem
A generic prompt produces generic marketing copy.
What I built
I built content-generation features — API, retrieval, storage and post-processing — inside an internal product that is still in development.
  • Python
  • FastAPI
  • RAG
  • LLMs
  • Supabase
  • MongoDB
logs → index → alert
Illustration

Raw logs

  • 12:04:31INFOauthlogin success user=•••
  • 12:04:31WARNauthlogin failed user=•••
  • 12:04:32ERRORauthtoken rejected
  • 12:04:32INFOauthlogout user=•••
  • 12:04:33INFOauthsession refresh
  • 12:04:33WARNauthlogin failed user=•••
  • 12:04:34INFOauthlogin success user=•••
  • 12:04:35ERRORauthmfa challenge timeout
  • 12:04:31INFOauthlogin success user=•••
  • 12:04:31WARNauthlogin failed user=•••
  • 12:04:32ERRORauthtoken rejected
  • 12:04:32INFOauthlogout user=•••
  • 12:04:33INFOauthsession refresh
  • 12:04:33WARNauthlogin failed user=•••
  • 12:04:34INFOauthlogin success user=•••
  • 12:04:35ERRORauthmfa challenge timeout

Structured event

{
  "level": "warn",
  "service": "auth",
  "event": "login_failed",
  "reason": "invalid_credentials"
}

Alert rule

WHEN
login_failed
RATE
above threshold
THEN
flag in the dashboard

Evaluating

Illustration of raw authentication log lines being parsed into structured documents and reviewed in a dashboard.

02Data platform · End-of-studies

Completed project

Authentication log analytics

Problem
Authentication logs stay unstructured and hard to search or review.
What I built
I built the ingest, index, analysis and dashboard path for authentication logs, as a completed end-of-studies project.
  • Logstash
  • Elasticsearch
  • Spark
  • Kibana

03Retrieval · pgvector

POC

Document question-answering pipeline

A language model cannot answer from documents it has never seen.

I built a document RAG proof of concept — chunking, vectors, retrieval and a grounded answer — that was not taken to deployment.

Proof of concept. Not pursued to deployment.

Read the Document question-answering pipeline case study

04Health AI · Clinical training

MVP in development

Clinical training simulator

Problem
An open chat cannot run a clinical case with controlled steps and consistent feedback.
What I built
I built functional modules of a clinical training MVP — orchestrator, state machine, evaluation and AI feedback — that is still in development.
  • Orchestration
  • State machine
  • Clinical engine
  • LLMs
clinical-sim / session loop
Illustration
Illustration of a training session loop: student action, state transition, clinical evaluation and AI feedback, producing a score.

04Inside the systems

How data becomes a product.

The architecture behind each project, step by step. Select a node, or use the arrow keys, to see what each part is responsible for.

An internal product in development: AI features that turn a brief into generated marketing content, grounded in retrieved context.

Full AI content generation platform case study →

  1. A user describes the content they need.

  2. API layer that receives requests and coordinates each step of generation.

    • Python
    • FastAPI

  3. Retrieves relevant context and prompts a language model with it.

    • RAG
    • LLM

  4. Persists application data, context and generated results.

    • Supabase
    • MongoDB

  5. Additional AI services applied to the generated material.

  6. Finished content returned to the user.

05How I work

From question to running system.

  1. 01

    Discovery

    Understand the problem, the people involved and the data that exists before choosing any tools.

    What comes out

    • Problem framing
    • Data inventory
    • Success criteria

  2. 02

    Architecture

    Design the flow end to end: sources, processing, storage, intelligence and the interfaces on top.

    What comes out

    • System diagram
    • Data model
    • Explicit trade-offs

  3. 03

    Build

    Ship in thin vertical slices, so pipeline, API and interface grow together and stay testable.

    What comes out

    • Pipelines & services
    • Typed APIs
    • Working interface

  4. 04

    Validation

    Check data quality, retrieval and model behaviour, and how the system fails, not just how it succeeds.

    What comes out

    • Test cases
    • Evaluation checks
    • Failure handling

  5. 05

    Delivery

    Deploy, observe and hand over something a team can understand and operate.

    What comes out

    • Deployment
    • Dashboards & alerts
    • Documentation

—Principles

  1. 01

    AI is a system, not a prompt.

    Models sit inside data, evaluation, serving and monitoring. I design the whole loop.

  2. 02

    Data quality comes before intelligence.

    Retrieval and outcomes are only as good as the data and structure underneath.

  3. 03

    Automation should remove friction.

    Pipelines and workflows should make teams faster, not create new manual work.

  4. 04

    Production matters more than prototypes.

    Latency, cost, failure modes and observability decide whether it is a demo or a product.

06Technology

A stack, organised by what it does.

Tools grouped by their role in a system, from the layer people touch down to where data lives. Press a project — or a tool — to see how its stack fits together.

20 tools across the stack

  1. 01

    Interfaces & APIs

    Where systems meet people and other software.

    • Power BIReporting skill
    • TableauReporting skill
    • Next.js · ReactThis site

  2. 02

    Intelligence

    Models and retrieval that turn data into answers and feedback.

  3. 03

    Orchestration

    Control flow that keeps multi-step processes predictable.

  4. 04

    Processing

    Pipelines that clean, transform and analyse data.

  5. 05

    Data & storage

    Stores chosen for how the data is queried.

  • Languages:
  • Python
  • SQL
  • TypeScript

07Contact

Have a system worth building?

I'm open to Data & AI engineering roles and project work: data platforms, RAG and LLM features, automation and full-stack products. Tell me about the problem.