00THE SIGNAL

Ibrahim Krayem · Generative AI engineer: LLM agents, RAG and document intelligence.

Turn complex information into decisions teams can use.

I design and build LLM applications, RAG systems and document workflows—from ingestion to evaluation and the user interface.

Experienced generative AI developer at TALAN, on Bpifrance missions since October 2024. PhD in Computer Science.

6+
years of experience
2+
years in generative AI

01THE PROBLEM

Start with the decision—not the model.

You do not need one more AI demo. You need a system that understands the right sources, exposes its reasoning path and produces a useful deliverable.

My research background brings disciplined modelling and evaluation. My engineering practice turns that discipline into software teams can operate.

  1. 01Clarify
  2. 02Connect
  3. 03Verify

02SIGNATURE CASE

From financial documents to a traceable SWOT workflow.

01 / SVI

Bpifrance · SVI agent · via TALAN

Bpifrance mission: a conversational financial agent for investment and holdings monitoring.

Problem
Retrieve financial data, analyse financial documents and produce usable deliverables in conversation, while keeping history and context across exchanges.
My role
Generative AI developer of an agile team of six, I build the agent and its tools.

Intervention — multi-format ingestion, sourced SWOT, SSE streaming and PowerPoint export.

  1. Orchestrate the agent with LangGraph and Claude, managing conversation history and context.
  2. Retrieve financial data through an MCP server.
  3. Build the tools: SWOT analysis with PowerPoint export, data visualisations, financial document analysis including capitalisation tables.
  4. Benchmark open-source models (Qwen) for extraction with Docling, and stream answers in real time over SSE.
Delivered transformationPDF · DATA → SWOT → PPTX

Agile team of 6 · production release coming soon.

Delivered result

An end-to-end conversational agent combining financial data, documents and deliverables, with its production release coming soon.

Limits and evidence — Usage and performance indicators belong to the client and are not published here. I can walk through the context, architecture choices and results in an interview.

  • LangGraph
  • Claude
  • MCP
  • LLM agents
  • Python
  • Docling
  • Qwen
  • SSE
  • Angular
  • PowerPoint
Explore all projects

SELECTED PROJECTS

Proof before promises.

For each case: the context, my contribution and what was delivered.
02 · DOCUMENT INTELLIGENCE

Making documents more useful.

ContextBpifrance · PICXEL · via TALAN

My roleWithin a SAFe organisation of about 50 developers, I worked on two tracks: a document analysis tool that extracts key information depending on the document type (signatures, dates, SIREN/SIRET checks…), and the rebuild of the search bar from strict matching to semantic search with embeddings.

Delivered resultBoth the extraction tool and the semantic search are in production.

03 · DOCTORAL RESEARCH

Understanding performance, faster.

ContextIRISA · TARAN research team

My roleDuring my PhD, I designed a queueing-theory model in Python that evaluates the latency of hybrid electrical and wireless networks on chip. As an R&D engineer, I then developed a trace-windowing method (window splitting and merging) that makes performance forecasts more accurate at lower computational cost.

Delivered resultAn analytical model up to 500× faster than the cycle-accurate Noxim simulator, with an error below 5%.

PERSONAL PROJECTS

Four things I built and run on my own, from idea to production.

Aqarih

Real-estate platform with AI search

Bilingual real-estate platform with an AI search assistant (Next.js, PostgreSQL, OpenAI API).

aqarih.com ↗ (new tab)

CalAL

Business app for aluminium workshops

Android and web app for aluminium workshops: calculations, panel cutting, stock and invoices; in production (Kotlin, Ktor, PostgreSQL).

calal.ikrayem.com ↗ (new tab)

NewsAI

LLM-written news digests

Arabic news digests written by LLMs and published on Telegram and WhatsApp (Python, OpenAI API, embeddings).

Personal server

Self-hosting my projects

Self-hosted Ubuntu server running CalAL, NewsAI and this portfolio: Caddy, PostgreSQL, scripted deployments, encrypted backups.

03THE SYSTEM

Build the full chain. Keep every step observable.

01

Frame

Decision, users, sources, risks and success criteria.

02

Build

Retrieval, agents, document processing, APIs and interfaces.

03

Observe

Tracing, sources, execution steps and operational signals.

04

Evaluate

Quality checks, limits, feedback and controlled iteration.

See methods and tools

EXPERTISE

Engineering useful, observable generative AI systems.

Technical capabilities translated into what they enable for a product and its users.

I connect models, data, interfaces and quality control so AI produces a usable result—not merely a plausible answer.

01

Generative AI

  • LLMs · RAG · LLM agents
  • MCP · tool calling
  • Claude · OpenAI API · Qwen
  • LangGraph · LangChain
  • Observability: Langfuse
  • Embeddings · Milvus · FAISS
  • Document extraction: Docling
02

Software engineering

  • Python
  • TypeScript · React · Next.js
  • Angular · Java
  • REST APIs · SSE
  • PostgreSQL · MongoDB · SQL
03

Cloud & delivery

  • Azure · AWS · Vercel
  • CI/CD · Jenkins
  • Datadog
  • Git · GitHub · GitLab
  • Agile · Scrum
04

Research

  • Performance modelling
  • Queueing theory
  • Pandas · SciPy
  • Scientific writing & teaching

FAQ

Frequently asked questions

What type of role are you looking for?

A generative AI engineering role involving LLM applications, RAG, agents, document intelligence and production concerns.

Can you work beyond the model layer?

Yes. The documented projects include Python services, Angular and React interfaces, APIs, databases, document pipelines, cloud and CI/CD tooling.

How do you handle LLM reliability?

By grounding work in explicit sources, exposing execution steps, adding observability and evaluating outputs against defined criteria. The exact evaluation protocol depends on the use case.

04TRAJECTORY

Research rigour, applied to production AI.

The same question runs through every stage: how can a complex system be made measurable, understandable and useful?

01

Experienced generative AI developer · TALAN

SVI financial agent (LangGraph, MCP) and PICXEL document intelligence for Bpifrance.

02

R&D engineer · IRISA

Continued the thesis: trace windowing for faster, more accurate network-on-chip performance analysis.

03

Doctoral researcher · IRISA

A queueing-theory model up to 500× faster than the Noxim simulator, with under 5% error.

04

PhD in Computer Science · University of Rennes

Research foundations in modelling, measurement and critical evaluation.

500×

Maximum speedup of my doctoral model over the Noxim simulator, with under 5% error.

Read the full journey

PROFESSIONAL JOURNEY

From system performance to generative AI.

From modelling hardware performance to delivering generative AI in production: the thread behind each step.
TALAN · Paris, France

Experienced Generative AI Developer

Generative AI developer on Bpifrance missions: the SVI conversational financial agent and the PICXEL document platform.

  • SVI (Bpifrance) — generative AI developer in an agile team of six: conversational financial agent built with LangGraph and Claude, managing conversation history and context; production release coming soon.
  • Connect the agent to financial data through an MCP server and build its tools: SWOT analysis with PowerPoint export, data visualisations, financial document analysis (including capitalisation tables).
  • Benchmarked open-source models (Qwen) against Claude models for document extraction with Docling; integrated real-time SSE streaming in the Angular frontend.
  • PICXEL (Bpifrance, SAFe organisation of ~50 developers) — built a document analysis tool that extracts key information by document type (signatures, dates, SIREN/SIRET checks…), in production.
  • Rebuilt the PICXEL search bar as embedding-based semantic search instead of strict matching, in production.
  • Other TALAN projects: RAG systems (retrieval part) and vector search with FAISS.
IRISA · TARAN · Lannion, France

Research & Development Engineer

Continued my doctoral research on the same theme: faster and more accurate performance analysis of networks on chip.

  • Developed a method that speeds up network-on-chip performance analysis by windowing real application traces (window splitting and merging, Poisson modelling).
  • Improved the accuracy of performance forecasts while reducing computational complexity, validated across several traffic types.
  • Implemented the trace processing pipeline in Python (Pandas, SciPy).
IRISA · TARAN · Lannion, France

Doctoral Researcher

Research into fast analytical methods for evaluating network-on-chip performance.

  • Thesis: “Methods for fast exploration of manycore architectures based network-on-chip with emerging technologies” (University of Rennes, defended February 2024).
  • Designed a queueing-theory model that evaluates the latency of hybrid electrical and wireless networks on chip: up to 500× faster than the cycle-accurate Noxim simulator, with an error below 5%.
  • Presented the research at national and international conferences.

TEACHING

Teaching alongside.

UPEC · EPISEN · Paris, France

Lecturer — Networks & Virtualisation

Teaching computer networks and virtualisation to engineering students.

  • Delivering lectures, tutorials and practical sessions in networks and virtualisation (VMware vSphere).
ENSSAT · University of Rennes · Lannion, France

Lecturer — Microcontrollers

Teaching microcontrollers to engineering students at ENSSAT.

  • Delivered tutorials and practical sessions in microcontrollers and supervised an interfacing project.

EDUCATION AND LANGUAGES

The foundations of the journey.

PhD in Computer Science

University of Rennes · Lannion, France

Engineering degree in Computer Science & Telecommunications

ENSTA Bretagne · Brest, France

Bachelor's degree in Computer Networks & Telecommunications Engineering

Lebanese University · Faculty of Technology · Saida, Lebanon

FrenchFluentEnglishProfessionalArabicNative
CERTIFICATIONS

Certifications

  • Microsoft Azure AI-900
  • Professional Scrum Master I
  • GitHub Copilot Certification
HOBBIES

Hobbies

  • Running
  • Walking
Follow my activities on Strava ↗
ENGAGEMENTS

Engagements

  • IEEE member
  • Order of Engineers - Lebanon
  • Volunteer with the French Red Cross

05NEXT SIGNAL

A generative AI role or mission to make concrete?

I bring a research mindset, end-to-end engineering and a focus on traceable outputs to generative AI products.

PriorityA generative AI engineering role. I am also open to a short, clearly scoped mission.

Portrait of Ibrahim Krayem

Paris · Fluent French · Professional English · Native Arabic