Senior Software Engineer · GenAI & Agentic Systems

Hi, I’m Akram Belajouza

I’m a Tech Lead and Full-Stack engineer with 15+ years of experience building scalable PHP systems and leading development teams. I also integrate GenAI, RAG, and automation into practical products using LLM APIs, Python, and modern workflow tools.

LinkedIn GitHub
Portrait of Akram Belajouza
Based in HamburgWorking with teams remotely

Expertise / 02

Full-stack foundations.
AI-focused direction.

Two complementary areas that define how I work today.

01
15+ years of experience

Backend / Full-Stack

  • Full-stack developmentPHP, Laravel, Symfony, CakePHP, SlimPHP, JavaScript, and jQuery
  • Data & persistenceMySQL, MariaDB, PostgreSQL, and SQL database design
  • Architecture & integrationSoftware and system architecture, API design, and third-party integration
  • Leadership & deliveryScrum, Kanban, team leadership, mentoring, and code review
02
Primary focus

GenAI / Automation

  • LLM integrationOpenAI, Gemini, and local LLM APIs; prompting, tool and function calling, and structured JSON outputs
  • RAG & semantic searchPython, embeddings, chunking, retrieval, LangChain, LangGraph, LangSmith, pgvector, and ChromaDB
  • Automation workflowsn8n, Zapier, Make, webhooks, and API orchestration
  • Voice & video pipelinesElevenLabs conversational agents, Tavus, and Heygen video avatars
  • AI-assisted developmentClaude Code, Codex, Cursor, Kilo Code, and OpenHands

AI projects / 03

Applied AI projects.

Architecture, trade-offs, and evidence—not just demos.

agentic-consensus.belajouza.com Workflow preview

Selected projects / 04

Systems I helped bring to life

Selected work across healthcare, fundraising, and identity management.

Healthcare · 202301

folioNet SSO

Unified authentication across multiple medical systems through a secure API and central messaging architecture.

  • PHP
  • Laravel
  • SSO
  • API
Healthcare · 2022–2302

ethikPool DICOM Viewer

Accelerated CT and MRI loading from minutes to seconds and reduced search times by 40% across 30+ client installations.

  • DICOM
  • PHP
  • Search
Healthcare · 2021–2203

painPool / my-Pain

Delivered an HL7 module to five practices within one month and introduced OAuth2-based Exchange integration.

  • HL7
  • OAuth2
  • REST
Fundraising · 2014–2104

Bildungsspender Platform

Developed a fundraising platform supporting more than €12M in donations and over 13,000 institutions.

  • PHP
  • Smarty
  • MySQL
  • Payments
Fundraising05

Donation Browser Extensions

Built Chrome, Firefox, and Edge extensions that brought direct donation integration into users’ shopping journeys.

  • JavaScript
  • Browser APIs
  • REST
Digital identity · 201306

VR-Networld SSO

Created a central portal for cooperative banks, integrating 10+ partner services with encryption and digital signatures.

  • SAML 2.0
  • OpenID
  • Security

Experience / 05

Built across healthcare, fundraising, and digital identity.

View selected work
2021 — 2026Remote

PHP Expert & Team Lead

UHB GmbH · Medical documentation systems

Led projects from conception to go-live, implemented secure PHP 8 and Laravel solutions, integrated HL7 and REST interfaces, administered client environments, and mentored the team.

Selected workfolioNet SSO · ethikPool DICOM viewer · painPool / my-Pain HL7 module
2014 — 2021Remote

Senior Web Developer

Bildungsspender gGmbH · Non-profit fundraising

Developed and maintained the platform from concept to production, integrating payment and partner APIs, improving performance, running server environments, and establishing development best practices.

Selected workBrowser extensions · API-powered shopping search · automated donation statistics · interactive institution map
2008 — 2013Remote

Team Lead

IT Future AG · Security & identity management

Planned and delivered security-critical applications, authentication platforms, and partner integrations using SAML 2.0, OpenID, OAuth2, encryption, and digital signatures.

Selected workVR-Networld Single Sign-On · Cloud-ID

Qualifications / 06

Education, learning & languages

The foundations and ongoing learning behind my work.

Education · 2010

Bachelor of Science

Technische Universität Darmstadt

Certifications · 2026

Applied GenAI

  • Introduction to LangChain — Python
  • LangGraph Essentials — Python
Languages

Multilingual

  • GermanFluent
  • EnglishFluent
  • FrenchFluent

Contact / 07

Let’s start a conversation.

Have a question, an opportunity, or an idea to discuss? Send me a message and I’ll get back to you.

Your details are used only to reply to this message.

AI engineering case study · Experimental

Agentic Consensus

I built an experimental framework to compare multi-agent review architectures and measure whether moderation and review loops improve results enough to justify their additional cost and latency.

  • Python
  • LangGraph
  • LangChain
  • FastAPI
  • OpenRouter
  • LangSmith
12-second workflow previewPlaceholder animation
Question→Intake→Author→Reviewer↺Final result

Question → fixed criteria → draft → review and revision → final synthesis

01 · What I built

An experiment, not just another AI demo.

The application runs the same task through different author-reviewer workflows and model combinations. It records quality, revision rounds, token use, provider-reported cost, and latency so the architectures can be compared rather than judged by intuition alone.

02 · Architectures

Three review strategies, one controlled comparison.

V1 · Baseline

Post-hoc review

Author → Reviewer ↺

The author writes first. The reviewer then derives criteria and requests revisions. It uses fewer calls, but its criteria may be influenced by the proposal.

V2 · Moderated

Fixed criteria first

Intake → Author ↔ Reviewer

A moderator clarifies the task and fixes acceptance criteria before writing begins, followed by author-reviewer loops and final synthesis.

V3 · Adversarial

Defect-oriented review

Intake → Author ↔ Critic

V2’s topology with a reviewer that actively searches for blocking defects, risks, ambiguities, and missing requirements.

03 · Technical interest

Designed for measurable, inspectable orchestration.

  • Graph orchestrationStateful author-reviewer loops built with LangGraph
  • Configurable modelsAnthropic and OpenAI models routed through OpenRouter
  • Structured verdictsPydantic schemas instead of parsing free-form reviewer text
  • Deterministic routingConditional edges decide approval, revision, stalling, or termination
  • Evaluation & tracingTokens, cost, latency, iterations, and LangSmith observability
  • Application layerFastAPI, SQLite persistence, Server-Sent Events, and automated tests

04 · Preliminary findings

Honest results before confident claims.

V1 uses fewer model calls. V2 and V3 introduce stronger evaluation controls by defining criteria independently before the answer is written.

V2 and V3 enable a fair reviewer comparison. Their shared topology isolates the difference between normal and adversarial reviewing.

Current statusThe first independent evaluation failed. More successful experiments are required before claiming a quality winner.