[ lab notes — 2026.10.21 ]

Samir G.

Hey, I'm Samir

AI engineer & full‑stack developer.

AI & agentic engineer at BeLogic AI · researcher at FEMTO-ST Institute · final year at ENSIA

I build and ship web and mobile products — e-commerce, management and institutional apps — and bring AI in where it actually helps. My research is on medical RAG, reinforcement learning for RAG and long-context LLMs, after earlier work on QPNN and MoQAdam.

index card

web: building since 2022
mobile: building since 2024
now: AI agents · BeLogic AI, Brussels
research: Medical RAG · FEMTO-ST, Besançon
open to: research · internships · engineering projects

Schedule a call CV GitHub

2026-10-21 · in progress

What I'm building right now

VOXORA — BeLogic AI

Voice and chat AI agents that answer a business’s customers and assist its whole team.
work · Brussels, Belgium
Industrial agentic system for healthcare

Grounded clinical answers — adaptive agent questioning and reinforcement learning for RAG.
lab · Besançon, France
Deepfake AI detection

Detecting AI-generated deepfake content, with a professor at Montreal University.
research · Montréal
TEXA — Ministry of Tourism

Multi-agent platform helping tourists discover Algeria’s cultural and natural sites.
production · Algeria
Filand — decision support

Agents that help companies make internal and external strategic decisions.
enterprise

Featured Engineering & Research

Flagship Systems & Deployed Platforms

Production-grade systems and applied research I have architected and shipped. Each deployment tackles high-stakes domain challenges—from nationwide Algerian government infrastructure to clinical decision architectures in France.

part A

Built at work

My current role at BeLogic AI in Brussels: an agent platform that talks with a business’s customers and works for its team.

VOXORA home page: ‘The phone rings. Who answers?’ with the AI agent answering an incoming call in German — **Visit the website** The home page, and the front agent answering in the chat on a phone (sample reply)

Chat with the VOXORA AI agent on a phone, the reply appearing word by word — **Visit the website** The home page, and the front agent answering in the chat on a phone (sample reply)

At work · BeLogic AI, Brussels

VOXORA // Voice & Chat AI Agents for European Businesses

Working MVP in development · website live · Oct 2026 – now

in one line One platform, two AI agents: a front agent that answers a business’s customers by chat and live voice, books into the real calendar and hands over to people, and a team agent that works for the staff and acts only after a person approves.

Small and medium European businesses — clinics, restaurants, hotels, shops — miss calls and messages when the office is closed or the team is busy, and every missed contact can be a lost booking. At BeLogic AI I build VOXORA. The front agent answers customers in their language from the company’s own documents, finds a free slot, holds it, reads it back and confirms the booking, or arranges a call-back; it says it is an AI at the start and never gives medical advice. The team agent prepares follow-ups, keeps calendars straight and answers staff questions; it proposes, and a person approves. Today it is a working MVP: text chat runs end to end with a real model and the website is live, and phone calls, the staff console and the team agent come in the next releases.

Field Note // What the First Live Test Revealed

"With a real model, the clinic agent could not book: it did not know today’s date or the clinic’s time zone. I added a rule-based date tool for six languages (243/243 on its test set), told the agent the local date and time, and let the slot search take an exact time. First answers took about 7 seconds, so embeddings moved to the background and search now waits at most 1.5 seconds before answering from full text."

▸ One Graph for Every Business: the agent is a single LangGraph graph; persona, rules and tools come from a versioned configuration, so a change is a publish, not a deployment, and models are reached only through an internal LLM gateway.
▸ Tools as Data: every tool is a versioned manifest that a gateway runs only if it is bound to the agent, behind an allowlist and a kill switch — defence in depth against prompt injection.
▸ 26 Microservices, Isolated by Tenant: Python and FastAPI, each with its own PostgreSQL database; identity only from signed tokens, forced row-level security, and health records in their own FHIR-aligned service.
▸ Requirements First, Every Change Gated: a 14-chapter requirements specification that every feature traces to; 1,553 unit, 308 integration and 158 website tests with a ratchet, plus type checks and security scanners.

Python · FastAPI LangGraph PostgreSQL LiveKit Astro · TypeScript Docker

Visit Live Site

how it works One conversation, through the platform

Customerwebsite chat or live voice in the browser
channel-servicewidget sessions
agent-orchestratorone LangGraph graph for every business
llm-gateway · agent-tool-gatewaymodel routing and metering · allowlist and kill switch
knowledge · scheduling · handover · tenantanswers with sources, slots and bookings, call-back tasks, local dates

The request path Every action goes through the tool gateway, even inside the platform.

The team agent’s capabilities: follow-ups and a daily digest, calendar and free time, answers with their sources, data questions with charts — **The team agent** For the whole staff, by text or voice; it proposes and a person approves. Planned for a later release.

part B

Built for institutions & research

Two ministries, a research lab and a robotics R&D agent: systems that had to be right before they could be fast.

https://texadz-tau.vercel.app/

Live on Vercel · 58 Wilayas

Launch Live Platform

Algerian Ministry of Tourism

Texa // National Tourism Intelligence

Production Web & Agent Platform · Algiers, Algeria

in one line A multi-agent trip planner for Algeria’s 58 wilayas that only suggests real routes and verified monuments, from the Ministry’s own data.

Commissioned to architect a unified digital tourism platform covering Algeria's 58 wilayas. Traditional travel assistants frequently hallucinate non-existent desert routes and closed landmarks. I engineered Texa as a multi-agent system where autonomous agents collaborate to synthesize personalized itineraries, verify historic monuments, and map realistic transit distances grounded directly in official Ministry datasets.

Field Note // The Engineering Challenge

"Standard travel LLMs hallucinate fictional desert oasis routes and invent fictitious transit connections. I constrained the multi-agent swarm with a strict GeoJSON transit topology and verified Ministry heritage records—ensuring zero hallucinated routes."

▸ Agent Swarm: Coordinated agents handling itinerary synthesis, cultural heritage data, and budget optimization.
▸ Spatial Grounding: Strict GeoJSON mapping across national monuments with zero-hallucination route calculation.
▸ Full-Stack Delivery: Interactive Next.js & React application with territory maps and official price registries in DA.

Next.js 14 React 18 TypeScript FastAPI Multi-Agent Systems GeoJSON Tailwind CSS

Visit Live Platform Architecture Overview GitHub

https://medical-assistant-dev-fdx4dfh5aq-uc.a.run.app/auth/

Google Cloud Run · Besançon Lab

Launch Clinical Assistant

Besançon Research Laboratory

MedRAG Ultra // Clinical Decision Support

RL for RAG & Active Probing · Besançon, France

in one line A clinical assistant that asks the doctor for the missing detail before it answers, trained with reinforcement learning.

In clinical consultation, an AI system that answers immediately when symptom descriptions are ambiguous is hazardous. Working in research collaboration with a clinical laboratory in Besançon, France, I engineered MedRAG Ultra with an active diagnostic questioning policy trained via Reinforcement Learning. When presented with ambiguous patient cases, the model proactively probes clinicians for critical missing details before synthesizing recommendations, backed by verbatim medical citations.

Field Note // Clinical Safety Breakthrough

"In clinical diagnostics, answering immediately when symptoms are ambiguous is dangerous. In research collaboration with the Besançon lab, we trained an active questioning RL policy that prompts clinicians for clarifying nuance before retrieval—reducing diagnostic uncertainty by over 54%."

▸ Active Probing: Reinforcement Learning agent executing diagnostic questioning policies to minimize clinical ambiguity.
▸ RL for RAG Core: Policy-guided retrieval trained with clinician reward modeling to suppress hallucinations.
▸ Verbatim Provenance: Sentence-level citation tracking grounded directly in peer-reviewed clinical ontologies.

Reinforcement Learning RL for RAG PyTorch Medical LLMs FastAPI Cloud Run Vector Databases

Visit Live Assistant Research Notes Private Repo

https://ramy-web-five.vercel.app/#/

Live on Vercel · 12ms Reflex Loop

Launch Live Platform

Autonomous Robotics & Embodied AI

RAMY // Intelligent Agent & Robotics

Embodied Planning & State Estimation · Live Vercel Deployment

in one line A robot planner with a 12 ms local reflex loop, so it dodges obstacles while bigger goals are planned in the background.

Autonomous robotic planning architecture engineered for dynamic physical environments. Decomposes complex operational goals into multi-horizon hierarchies, separating reactive low-latency obstacle reflexes from high-level state estimation and closed-loop telemetry.

Field Note // Real-Time Control Breakthrough

"Cloud inference is too slow for physical robotics. RAMY pairs a fast 12ms local reactive reflex loop with asynchronous high-level hierarchical goal decomposition."

▸ Dual-Horizon Planning: Decouples low-latency 12ms reactive obstacle reflexes from high-level goal decomposition.
▸ State Estimation: Multi-sensor fusion and state estimation pipeline operating reliably under dynamic physical constraints.
▸ Live Telemetry Console: Real-time streaming dashboard visualizing sensory maps and control states in real time.

Autonomous Agents Robotics Python State Estimation Reinforcement Learning Telemetry Vercel

Visit Live Platform Architecture Overview Architecture Specifications

0.856 Faithfulness

SPIRAL-RAG // Regulatory Archive Intelligence

Ministry of Higher Education Regulatory Corpus

in one line Ask about Algerian higher-education law in Arabic, French or Darija and get an answer that quotes the official text: 0.856 faithfulness on 330 questions.

Multilingual statutory retrieval pipeline built for Algerian Higher Education legal archives. Seamlessly synthesizes cross-referenced decrees across Arabic, administrative French, and colloquial Darija, validated against a 330-question benchmark with 0.856 Faithfulness.

Benchmark Result // Rigorous Verification

"Algerian legal corpora mix Modern Standard Arabic decrees, French administrative codices, and Darija phrasing. Evaluated against 330 official questions to achieve an empirical 0.856 Faithfulness benchmark."

▸ Multilingual Normalization: Unified embeddings across Arabic, administrative French, and Darija legal phrasing.
▸ Triad Faithfulness Benchmark: Empirically scored across 330 official legal questions with zero hallucinations.
▸ Statutory Citation Tracking: Grounded sentence-level provenance linking answers directly to ministerial gazettes.

Python RAG Triad Gemini Pro Llama 3.3 FAISS Legal NLP

View GitHub Repository Evaluation Notes

part C

Built for clients, freelance

Algerian businesses that sell, deliver and care for people every day: shops, a home-care network and an operations platform, each built end to end.

SkinDZ homepage on desktop: a woman in a hijab, the headline and a scrolling strip of brands — **Visit the live shop** The homepage, the same page in Arabic, and a product page with the 30-second order form

SkinDZ homepage in Arabic with a right-to-left layout — **Visit the live shop** The homepage, the same page in Arabic, and a product page with the 30-second order form

Freelance · Algerian beauty brand

SkinDZ // Cash-on-Delivery Beauty Commerce

Live e-commerce platform · Algeria · Oct 2026

in one line A trilingual beauty shop where Algerians order in 30 seconds and pay the courier, built alone for a brand with about 400K Facebook followers.

Designed and built, alone, the online shop of an Algerian skincare brand followed by about 400K people on Facebook — so it had to hold up on launch day. Customers order in 30 seconds and pay cash on delivery in all 69 wilayas, in French, Arabic (right-to-left) or English, and the owner runs everything from a back-office on her phone.

Field Note // The Engineering Challenge

"Algerian shoppers find products on Facebook, distrust online payment and pay the courier. So the price shown must be the price charged: one shipping calculator runs in the browser and on the server, and every total is recomputed server-side. Orders reach the owner's Google Sheet after the response, so a slow sheet never delays a checkout."

▸ COD Checkout: every wilaya and commune, home or stop-desk delivery, and the total with delivery shown before confirming.
▸ Skin Diagnostic: a short quiz that builds a morning and evening routine from the real catalogue, shareable on WhatsApp.
▸ Owner Back-Office: orders, products, homepage editor, delivery prices, comments, ad pixels and a maintenance switch, in three languages.
▸ Delivery: Docker on a VPS with Coolify, auto-deployed on every merge · 42 merged pull requests · 30 unit tests.

Next.js 16 React 19 TypeScript PostgreSQL Prisma Docker i18n · RTL

Visit Live Shop

Doctor Darna homepage in French: a watercolor of a doctor visiting a family at home, with the call and booking buttons — **Visit the live site** The homepage in French, the same page in Arabic, and the map of the wilayas served

Doctor Darna homepage in Arabic with a right-to-left layout — **Visit the live site** The homepage in French, the same page in Arabic, and the map of the wilayas served

Freelance · Home medical care, Algeria

Doctor Darna // Home Healthcare Platform

Live trilingual site & custom admin · Algiers & Sétif · Oct 2026

in one line A warm, trilingual home-care website with its own admin, ready to open a page for any of the 69 wilayas in minutes.

Doctor Darna sends doctors, nurses, physiotherapists and ambulances to patients’ homes, 24/7, from clinics in Algiers and Sétif. I rebuilt its aging, French-only WordPress site into a warm, trilingual Next.js platform — French, Arabic (right-to-left) and English — with a custom admin panel the team uses every day. I did all of it: design, front end, back end, admin, SEO, hosting and DevOps.

Field Note // The Engineering Challenge

"A clinic network cannot open 69 thin copies of one page without Google treating them as spam. Each wilaya is an admin record that can only go live with a local introduction, a real service and a municipality — then it gets its own page, sitemap entry and structured data, and Bing is notified at once. Opening a new wilaya takes minutes, with no code and no deployment."

▸ Built to Get a Call: a sticky call button, a phone tab bar, WhatsApp pre-filled with the service or wilaya, and a booking form that falls back to WhatsApp.
▸ Custom Admin, No CMS: edit any page with a live preview, add clinics from a Google Maps link, and handle requests from an inbox with CSV export — in French, Arabic or English.
▸ SEO & AI Visibility: schema.org for every clinic and service, hreflang, llms.txt, IndexNow and 190+ old WordPress URLs redirected; about 120 indexed pages.
▸ Quality & Ops: a WCAG 2.2 AA accessibility pass; CI, CodeQL and uptime checks every 15 minutes on GitHub Actions · 28 merged pull requests.

Next.js 16 React 19 TypeScript Tailwind CSS 4 libSQL / Turso GitHub Actions i18n · RTL

Visit Live Site

El Basma home page on a computer: two arch-framed photos of djellabas on a plum background, beside the slogan — **Visit the live shop** The home page, an ad landing page written by Gemini, and a product page

An AI-generated ad landing page in Arabic for one djellaba, on a phone — **Visit the live shop** The home page, an ad landing page written by Gemini, and a product page

Freelance · Djellaba boutique, Maghnia

El Basma // Storefront, Back-Office & AI Landing Pages

Live e-commerce · Tlemcen, Algeria · built 2025, rebuilt as v2 in Oct 2026

in one line A djellaba boutique’s whole online shop, rebuilt: a four-language storefront with one-tap cash-on-delivery ordering, a back-office to run it, and an AI that turns a product photo into an ad page.

El Basma (البصمة) is a women’s djellaba boutique in Maghnia. I first gave it a new identity, a one-page cash-on-delivery checkout and a Gemini-powered landing-page generator. Then I rebuilt the whole shop as v2: the owner could only edit a few texts from the website, the cart sent one order per item, and the prices came from the browser. Now a premium storefront in Arabic, Darija, French and English sits on a full back-office — orders, abandoned checkouts, products, delivery prices for 58 wilayas, reviews, ad pixels and a Google Sheets copy of every order — on an Express API shared with the team’s mobile app.

Field Note // The Engineering Challenge

"The owner wanted ad pages like Algerian COD tools generate, without designing each one. Gemini writes the page from the real product only — visible details and real shop policies, no invented reviews or urgency. When the client found campaign-style AI photos ‘too AI’, I made the image model shoot like an ordinary phone: worn in the shop, a hand on the buttons, a mirror selfie."

▸ Brand Identity: a hand-lettered Ruqaa logo where the dot of the ب becomes an aqad, the knotted djellaba button, with a gold thread; plum, rose and ivory, and arch photo frames inspired by Tlemcen.
▸ A Premium Storefront: a photo hero, best sellers, promotions and customer screenshots the owner turns on and off; a quick-order sheet on every product card, a sticky order bar on phones, and wilaya and commune pickers with a free-delivery meter.
▸ A Back-Office to Run the Shop: a Today page with the orders to call, an order workflow with one-tap call and WhatsApp, abandoned checkouts saved as soon as a phone number is typed, delivery prices for 58 wilayas, review moderation and a Google Sheets copy of every order.
▸ Orders the Server Can Trust: a cart is now one order with a readable number; prices, stock and delivery fees are computed on the server, and public forms have rate limits and honeypots. 162 automated backend checks pass, and every schema change kept the team’s app working.
▸ AI Landing Pages: a product photo becomes a ready-to-advertise page in four languages with 3 photos, reviewed in a live phone preview; the API key stays server-side.

React 18 Vite 5 Node.js / Express Turso / libSQL Google Gemini Google Sheets i18n · RTL

Visit Live Shop

v1 → v2 The whole shop, rebuilt

The first version of the home page: a light page with two mannequin photos and the slogan in Arabic — **Before → after** Drag across the home page to compare.

The rebuilt home page: two arch-framed photos of djellabas on a plum background, beside the slogan — **Before → after** Drag across the home page to compare.

The back-office Today page: orders to call, today’s and the month’s figures, a 14-day chart and best sellers — **The back-office** In Arabic, French or English, light or dark. Shown with test orders.

The orders list with its status tabs, search and bulk status change — **The back-office** In Arabic, French or English, light or dark. Shown with test orders.

part two · same client

El Basma App // The Team’s iOS & Android Back Office

Rebrand & upgrade · React Native / Expo · Oct 2026 · for the shop’s team, not on the stores

The staff app's dashboard before the rebrand: a generic look with default colours — **Before → after** Drag across the dashboard to compare.

**Watch the demo** The dashboard, a product, orders and delivery prices (test data).

in one line The app the boutique’s team uses for orders, stock and customer messages now wears the website’s hand-lettered identity, runs on Expo SDK 57, and opens again on iPhone.

The El Basma team runs the shop from a phone app: orders with one-tap call or WhatsApp, products with stock per colour and size, and a customer inbox with voice and video messages — in Arabic, Darija, French and English. It still had a generic look, and it was built on Expo SDK 50: the current Expo Go on the owner’s iPhone refused to open it. I rebranded it to match the website, upgraded the whole stack and fixed the bugs that blocked the team, with AI-assisted development.

Field Note // The Bug That Stopped the iPhone

"On iPhone the app died at start-up with ‘JSON Parse error: Unexpected character’. The remote config file carried an invisible byte-order mark, and iOS refused it. The loader now strips it and falls back to the next mirror on a bad file — and logging in finally goes to the server chosen in Settings, not always to production."

▸ Rebrand: the website’s hand-lettered logo, a new app icon and splash screen, and brand tokens applied across about 43 screens, in light and dark mode.
▸ Upgrade: Expo SDK 50 → 57, React Native 0.73 → 0.86, React 18 → 19.2; expo-av replaced with expo-audio and expo-video for voice and video messages.
▸ Verified: opens in the current Expo Go on iPhone; 23 of 23 tests pass (the suite could not run before); the iOS production bundle builds.

React Native 0.86 Expo SDK 57 React 19.2 TypeScript 6 i18next · RTL Jest

Dahaoui Boutique home page on a computer: two arched photos of djellabas on black satin, beside the slogan in gold — **Visit the live shop** The home page, and the home and a product page on a phone, in Arabic

Freelance · Women’s clothing shop, Maghnia

Dahaoui Boutique // Black-and-Gold Storefront & Back-Office

Live e-commerce · Tlemcen, Algeria · Oct 2026 · the El Basma platform, adapted to a second client

in one line A women’s traditional clothing shop rebuilt in black and gold from its own shop sign: a four-language storefront with one-tap cash-on-delivery ordering for the 69 wilayas, a full back-office, and a secured API that took over the live data without downtime.

Mahal Dahaoui (محل دحاوي) sells djellabas, caftans, abayas and takchitas in Maghnia. Its old site looked like a template — stacked promo bars, emoji headlines, a fake “people are watching” counter — took one product per order, and trusted the prices the browser sent. I adapted the platform I built for El Basma to a second client: a black-and-gold identity drawn from the shop’s round sign, a premium storefront in Arabic, Darija, French and English, a complete back-office that keeps the shop’s own tools (customers, returns, statistics, a team with roles), and the same API rebuilt on MongoDB.

Field Note // Switching Without Downtime

"The new API had to take over a live shop with its real products, orders and settings. Start-up migrations only add what is missing — wilaya codes and prices, settings carried over from the old ones — and every call the old website made was replayed against the new API before the switch. The new site went live on the existing data, with no downtime and no production configuration change."

▸ An Identity From the Shop Sign: a round black-and-gold badge, the «محل دحاوي» wordmark in gold Aref Ruqaa, and favicons, app icons and a social preview generated from the sign itself.
▸ A Storefront That Sells Honestly: the fake counter and emoji headlines are gone; arched photos on black satin, a quick-order sheet, a multi-item cart sent as one order, 69 wilayas with their communes, and free delivery everywhere by default.
▸ A Back-Office With Roles: orders with search, bulk status, CSV export and one-tap call or WhatsApp, abandoned orders, delivery prices, pixels and Google Sheets — plus customers, returns, statistics, and managers who only see what their job needs.
▸ A Secured API on MongoDB: prices, stock and fees computed on the server; public routes whitelisted, the public reset of delivery prices closed, unpublished products hidden, rate limits and honeypots. 362 automated checks pass.

React 18 Vite 5 Node.js / Express MongoDB / Mongoose Cloudinary Google Sheets i18n · RTL

Visit Live Shop

before → after From a template to a boutique

The old home page: stacked red and gold promo bars, emoji headlines and the phone number in the middle of the page — **Before → after** Drag across the home page to compare.

The back-office Today page: orders to call, today’s and the month’s figures and a 14-day chart — **The back-office** In Arabic, French or English, with roles for the team. Shown with a test database, without customer data.

Delivery settings: free delivery everywhere, with home and stop-desk prices for each of the 69 wilayas — **The back-office** In Arabic, French or English, with roles for the team. Shown with a test database, without customer data.

The Akkida home page in Arabic: the headline 'a small team, and no order is lost', the five people of the team drawn by hand, and 'a day in the life of an order' told in six comic panels with Darija speech bubbles — **The home page, told as a comic** A hand-drawn day of the team, from the order arriving in the sheet to the courier at the door. It scrolls by itself; hover to see it again. Below: the courier's run and today's confirmation list, on phones.

The courier's run on a phone: parcels grouped by area, with Call, Map, Delivered and Not delivered on every row — **The home page, told as a comic** A hand-drawn day of the team, from the order arriving in the sheet to the courier at the door. It scrolls by itself; hover to see it again. Below: the courier's run and today's confirmation list, on phones.

Freelance · Algerian cash-on-delivery business

Akkida // Order Confirmation & Delivery Operations

In production · login required · Algeria · since Oct 2026

in one line Every call becomes a confirmed order: an Arabic-first platform that runs a cash-on-delivery shop’s confirmation calls, shipping and couriers, with a live view for the owner.

In Algeria most online orders are paid cash on delivery, and each one must be confirmed by phone before it ships: an unconfirmed parcel comes back as a return, with the shipping paid twice. This business ran it on Google Sheets and phones. I built Akkida alone, from product analysis and brand to deployment — orders flow in from the sheets, agents call under strict time rules, parcels go to the delivery company or the shop’s own couriers, and the owner sees everything live, in Arabic first, with French and English.

Field Note // The Engineering Challenge

"A browser can’t know whether the customer picked up. So the tap on «اتصال» (Call) is the attempt: it is logged and timed at once, and the agent picks the result. That made the client’s rules enforceable without a native app — a first call within 15 minutes, 3 attempts a day 2 hours apart, a second day, then automatic cancellation, on a work calendar with shifts and breaks."

▸ Call Discipline Engine: a pure rules engine checked every minute; a warning at minute 10, missed deadlines recorded, and an automatic ban past the owner’s allowance.
▸ One-Handed Courier Run: parcels grouped by wilaya and commune, one tap per parcel, and “Did you collect 4 500 DA?” before a parcel closes for good.
▸ Integrations: the EcoTrack delivery API (bulk create, 30-minute status sync, labels), a Telegram bot for alerts, and Google Sheets through a signed webhook.
▸ Built to Hold: 14 modules with enforced boundaries, optimistic locking and an append-only order history · 409 automated tests · 64 merged pull requests.

Next.js 15 React 19 TypeScript PostgreSQL Drizzle ORM Vitest · Playwright Arabic RTL

Private client app · login required

ch. 03 — experience

Where I’ve worked

My current role, four internships, client work alongside study, and the teaching I do — newest first. The dates are in the margin.

Sep 2026 – now Voice & chat AI agents for European businesses.
AI & Agentic Engineer

BeLogic AI · Brussels, Belgium

current role

Building VOXORA, BeLogic AI’s agent platform: a front agent that answers a business’s customers by chat and live voice, books into the real calendar and hands over to people, and a team agent that assists the whole staff and acts only after a person approves.
Agent — One LangGraph graph for every business; behaviour from a versioned configuration, models only through an internal LLM gateway.
Platform — 26 Python and FastAPI microservices, each with its own PostgreSQL database and row-level security.
Quality — 1,553 unit, 308 integration and 158 website tests, with type checks and security scanners on every change.

Request path
Customer → channel → LangGraph orchestrator → tool gateway → knowledge, scheduling, handover
Status
Working MVP: text chat runs end to end with a real model and the website is live; phone calls, the staff console and the team agent come next.
Python · FastAPI · LangGraph · PostgreSQL · LiveKit · Astro / TypeScript
Read the case study The VOXORA website, with the front agent answering in the chat
Jun – Aug 2026 Real-time NL-to-SQL querying revenue without SQL.
Data Scientist Intern

Djezzy (DSIP Program) × majestEYE · Algiers, Algeria

in production

Architected and deployed an autonomous natural language-to-SQL business intelligence agent empowering non-technical enterprise experts to query real-time revenue, subscriber churn, and network telemetry without writing SQL.
Core Agent — LLM query translation with schema grounding and syntax validation.
Streaming & Scale — Event-driven Kafka architecture decoupling requests from warehouse load.
Business Impact — Self-serve pipeline eliminating manual analyst ticketing and bottlenecks.

Data Pipeline
User Prompt → LLM Schema Grounding → Apache Kafka Event Queue → SQL Data Warehouse → BI Dashboard
Decoupling
Asynchronous Kafka message streaming buffers telemetry surges and protects warehouse compute.
Python · LLMs / Agentic AI · Apache Kafka · NL-to-SQL · SQL Data Warehouse · Event-Driven Architecture
Watch the demo User types query in plain language → LLM generates validated SQL → Kafka streams telemetry & BI metrics
Dec 2025 – Feb 2026 Synchronized web storefront and mobile app with shared TypeScript models.
Full-Stack Software Engineer Intern

Algérie Poste · Algiers, Algeria

deployed

Developed an enterprise multi-platform e-commerce ecosystem consisting of a responsive web customer storefront and cross-platform native mobile application powered by shared TypeScript architectures and merchant intelligence.
Dual Clients — Synchronized React 18 web & Expo mobile app with shared TypeScript models.
Merchant Ops — Store management, catalog hierarchy, order fulfillment, and RBAC.
Real-Time Ops — Live merchant analytics dashboard with sales velocity and inventory alerts.

Client Layer
React 18 Web Storefront + React Native (Expo) Mobile App with shared TypeScript domain contracts
Platform Core
Unified REST API Gateway → Multi-tenant merchant stores, shopping carts, RBAC & live sales metrics
React 18 · React Native (Expo) · TypeScript · REST APIs · RBAC Security · Merchant Analytics
Watch the demo Cross-platform React Native & Web UI → Shared TypeScript contracts & REST gateway → Merchant RBAC & order telemetry
Jul – Aug 2025 Redesigned daisy-chained bus topology into a redundant star, cutting latency by 40%.
Network & Systems Engineer Intern

ENOF SPA (National Mining Enterprise) · Algiers & Industrial Mining Units, Algeria

Audited, mapped, and diagnosed enterprise local area networks (LAN) bridging ENOF corporate headquarters and 6 distributed industrial mining branches, isolating bottlenecks and engineering high-availability topologies.
Infra Audit — Enterprise LAN audit and traffic diagnostics across HQ and 6 mining sites.
Star Topology — Redundant star core topology eliminating single points of failure.
QoS & VLANs — Managed VLAN isolation for telemetry, corporate ERP, and VoIP QoS.

Topology
Replaced single-bus daisy chains with a redundant star core switch architecture across 6 mining branches
Segmentation
Managed VLAN separation between operational telemetry, corporate ERP, and VoIP traffic with QoS prioritization
Enterprise LAN · Redundant Star Topology · VLAN Segmentation · QoS Prioritization · Network Diagnostics · Industrial Hardware
Jul – Aug 2023 Rapid 2-week agile sprint iterations delivering clean modular UI.
Frontend Developer Intern

HR Technology · Algiers, Algeria

Developed client-side features and automated administrative workflows for enterprise human resource management software within rigorous 2-week agile sprint cycles.
Modular Client — Clean, modular UI components in vanilla JavaScript and semantic HTML5/CSS3.
REST Integration — Backend microservice integration for automated leave request workflows.
Agile Cadence — Bi-weekly sprints, reviews, and continuous code reviews with zero defects.

Interface
Modular vanilla JavaScript components consuming RESTful microservice endpoints for employee personnel records
Cadence
2-week sprint releases with daily standups, backlog grooming, peer code reviews, and zero regressions
JavaScript (ES6+) · REST API Integration · Agile / Scrum · UI Performance · Workflow Systems · HTML5 / CSS3
2023 – now alongside study
Client work, on the side

Production e-commerce platforms, distributed backends, webhook automations, and cross-platform mobile systems engineered for business clients.
Dahaoui Boutique

Freelance Full-Stack Engineer · 2025 – 2026 · Remote

A women’s traditional clothing shop in Maghnia: a black-and-gold storefront, a full back-office and a secured Express + MongoDB API.

Four-language storefront with one-tap cash-on-delivery ordering for 69 wilayas
Back-office with orders, customers, returns, statistics and team roles
362 automated API checks; switched to the new version with no downtime

React · Node.js · Express · MongoDB · Cloudinary

case study live shop (Dahaoui Boutique)
Self-Employed

Freelance Mobile Developer · 2023 – Present · Remote

Engineered 5+ commercial cross-platform mobile applications using Flutter for business clients.

Custom Flutter apps with reactive state management and 60fps UI
Offline-first local storage, push notifications, secure RESTful APIs
Full development lifecycle from wireframing to store release

Flutter · Dart · Firebase · REST APIs · Mobile UI/UX

5+ Apps Shipped

Teaching & coaching

2024
Competitive Programming Coach

Algerian Olympiad in Informatics (AOI) · National Program • Algeria
Curriculum & Algorithms — Coached students in advanced algorithms, graph theory, and dynamic programming.
Competition Strategy — Prepared and mentored students for national competitions with personalized strategies.
Problem Creation — Created comprehensive practice problems and structured training modules.
2023 – 2024
EIS

Algorithms Teaching Assistant

El Ijtihad School · Algiers, Algeria
Course Instruction — Instructed students in Algorithms 1 & 2, covering fundamental to advanced concepts.
Pedagogy Design — Designed curriculum bridging theoretical concepts with hands-on coding challenges.
Interactive Teaching — Enhanced student engagement and comprehension through interactive teaching methods.
2023
IT Manager & Technical Lead

Scientific Club (EBEC), ENSIA · Algiers, Algeria
Systems & Ops — Managed IT infrastructure and technical operations for student scientific club.
Workshops & Events — Coordinated technical workshops and events for 50+ club members.
Digital Platforms — Oversaw digital platforms and organization communication systems.

ch. 04 — research

Research

Three preprints, all first-author, each with its code public. Every paper starts with what it does in plain words; the maths and the code follow for those who want them.

QPNN-v2 QuantumPath Neural Networks Architecture Concept — **Fig. 1** QuantumPath multi-path inference with Orthogonal Interference Mixing O ∈ O(K) and Von Neumann Entropy.

Guenchi, Samir. "QuantumPath Neural Networks: Orthogonal Path Interference and Von Neumann Entropy for Principled Uncertainty Estimation." 2025.

code on GitHub

preprint · 2025 · Quantum Information Theory × Deep Learning · PyTorch ≥ 2.0 · Open Source

QPNN-v2: QuantumPath Neural Networks

Orthogonal Path Interference and Von Neumann Entropy for Principled Uncertainty Estimation

in plain wordsA model that can say “I’m not sure”. Ensembles of several networks give honest uncertainty but cost several times the compute; QPNN-v2 gets ensemble-style diversity from one network trained once, with about 72% fewer parameters than a four-model ensemble.

12.6M: Total Params (vs 44.8M)
K = 4 Paths: Complex Representations
Single Pass: O(1) Training Cost
OOD-AUROC: SVHN Detection via VNE

Uncertainty estimation in deep learning suffers from a classic trilemma: single CNN softmax confidences are uncalibrated; MC Dropout heuristics poorly approximate deep priors; and Deep Ensembles demand prohibitive O(M) training and memory footprint.

QPNN-v2 achieves ensemble-grade diversity in a single training pass through a quantum-inspired multi-path architecture where representations evolve in the complex domain ℂ^d, are mixed via a provably entropy-preserving orthogonal operator O ∈ O(K), and uncertainty is quantified by the Von Neumann entropy of the path Gram matrix.

What’s new in it

Complex Path Representations CPR
Each of K paths outputs h_k ∈ ℂ^d (magnitude r_k > 0 & phase θ_k ∈ ℝ), doubling neuron degrees of freedom without native complex layers. Amplitudes α_k = r_k·e^iθ_k strictly respect the Born rule: ∑_k |α̃_k|² = 1.
Orthogonal Interference Mixing OIM
Paths mixed via O = exp(M) ∈ O(K) with skew-symmetric generator M = (W − Wᵀ)/2. Guarantees exact OᵀO = I with only K(K−1)/2 parameters—no soft regularization penalty needed.
Von Neumann Entropy Uncertainty VNE
Density matrix ρ = G̃ / K from normalized Gram matrix G̃_ij = ⟨h_i, h_j⟩ / (‖h_i‖·‖h_j‖). S_vN(ρ) = −∑ λ_k log λ_k provides a continuous, permutation-invariant, and concave uncertainty score.

The guarantees, the code and the specs

Entropy Invariance (Proposition 1): S_vN(ρ) is strictly invariant under right-multiplication by any O ∈ O(K). Orthogonal mixing redistributes information across paths without destroying diversity.
Tight Information Bounds (Proposition 2): S_vN = 0 iff all K paths are identical (rank-1); S_vN = log K iff all paths are mutually orthogonal (G̃ = I). Training maximises S_vN as an information-theoretic regulariser.
72% Parameter Compression: Total parameter footprint is ~12.6M params (ResNet-18 backbone + 4 complex paths) vs. 44.8M params for a 4-member Deep Ensemble, delivering single-pass inference.

Quick Start & Factory Usage:

from qpnn_v2 import create_qpnn

# Presets: "fast" (K=2), "balanced" (K=4), "accurate" (K=6)
model = create_qpnn(
    input_dim=512, hidden_dim=256, output_dim=10,
    task="classification", preset="balanced"
)

out = model(x)             # Single forward pass: x (B, 512)
probs = model.predict_proba(x)  # (B, 10) class probabilities
vne = out["vne"]           # (B,) calibrated uncertainty score ∈ [0, 1]

Collapse Mechanism	Mathematical Formulation	Extra Parameters	Optimal Use Case
Born	p_k = \|α̃_k\|² (pure quantum collapse)	0 extra	Purest quantum information interpretation
Soft (Recommended)	Attention scorer + amplitude prior + learned τ	~H² / 2	Best expected calibration error (ECE)
Adaptive	Input-dependent τ ∈ [τ_min, τ_max]	~K·H + H	Highest variance out-of-distribution inputs

MoQAdam Optimization Landscape and Loss Surface Analysis — **Fig. 2** Objective function loss surface analysis: SGRF residual feedback automatically anneals near global and local minima.

Guenchi, Samir. "MoQAdam: Moment-Normalized Quantization Adam via SNR-Gated Residual Feedback and Decoupled Moment Updates." 2026.

code on GitHub

preprint · 2026 · Distributed Optimization & Compression · PyTorch DDP · NCCL Ready

MoQAdam: Moment-Normalized Quantization Adam

Distributed Training with Provably Bounded, Automatically Annealing Gradient Quantization

in plain wordsTraining on many GPUs is slowed down by the gradients the machines keep sending each other. MoQAdam is a drop-in replacement for the AdamW optimiser that compresses that traffic to 8 or 4 bits for under 3% extra compute, while keeping its convergence guarantees.

< 3%: Overhead vs AdamW
8-bit / 4-bit: Configurable VNQ
2.5P: State with fp16 Residual
Unbiased: Variance Estimator

Inter-node communication bandwidth is the primary bottleneck in distributed deep learning clusters. Standard gradient quantization methods fail either by introducing layer-dependent variance spikes or by inducing uncontrolled error-feedback oscillations near convergence.

MoQAdam serves as a drop-in AdamW replacement for distributed PyTorch DDP training. By quantizing gradients in Adam’s own normalized space, gating residual error feedback by the running signal-to-noise ratio, and decoupling moment updates, MoQAdam compresses communication while preserving theoretical convergence bounds.

What’s new in it

Variance-Normalized Quantization VNQ
Quantizes in Adam's own scale-invariant normalized space g_t / σ_t. The quantization error bound is mathematically layer-invariant and distribution-free, eliminating gradient scale disparities.
SNR-Gated Residual Feedback SGRF
Gates the error feedback residual by gradient SNR: γ_t = clamp(ρ̄_t / ρ_ref, 0, 1). The residual automatically vanishes near minima, preventing oscillations without manual decay schedules.
Decoupled Moment Updates DMU
Updates second moment v_t from exact raw gradients g_t, while updating first moment m_t from quantized q_t. Guarantees v̂_t remains an unbiased estimator of 𝔼[g_t²].

The guarantees, the code and the specs

PyTorch DDP Native Hook: MoQAdamDDPHook registers per-parameter math hooks + float16 AllReduce transport hooks automatically with PyTorch ≥ 2.1.
2.5P Memory Footprint: Halves residual memory using float16 residuals (sustained state: m + v + e_fp16 = 2.5P) with transient 1P AllReduce buffers freed immediately.
Dynamic Routing & MoE Ready: Handles skipped layers in Mixture of Experts routing (find_unused_parameters=True) by synthesizing zero-gradient v_t decay to maintain cluster state parity.

Multi-GPU Distributed Data Parallel (DDP) Integration:

from moqadam import MoQAdam
from moqadam_ddp import MoQAdamDDPHook

base_opt = MoQAdam(
    model.parameters(), lr=1e-3, quant_bits=8,
    snr_ref=0.10, residual_dtype=torch.float16
)
# Registers per-parameter math hook and fp16 transport comm hook
opt = MoQAdamDDPHook(base_opt, model)

for step, (x, y) in enumerate(loader):
    opt.zero_grad()
    loss = criterion(model(x), y)
    loss.backward()  # Math hook: VNQ + SGRF; Transport: fp16 AllReduce
    opt.step()

Optimizer Configuration	State Per Parameter	Memory vs AdamW	Target Workload
Vanilla AdamW	m + v = 2P	Baseline	Single GPU / unconstrained interconnect
MoQAdam (fp32 residual)	m + v + e = 3P	+1.0P	High-precision distributed clusters
MoQAdam (fp16 residual)	m + v + e_fp16 = 2.5P	+0.5P	Bandwidth-bottlenecked multi-node & federated DDP

SleepStagerNetV4 Single-Channel EEG Sleep Stage Neural Classification Architecture and Brainwave Dynamics — **Fig. 3** Polysomnographic neural architecture decoding 5 AASM stages (Wake, N1, N2, N3, REM) from single-channel 178 Hz EEG.

Guenchi Samir, Guerroudj Abdennour. "SleepStagerNetV4: Breaking the κ = 0.57 Barrier in Automated Single-Channel EEG Sleep Staging via Four Synergistic State-of-the-Art Innovations." 2026.

code on GitHub

clinical AI · 2026 · Biomedical Signal Processing & Time-Series · PyTorch · Tested on NVIDIA H100

SleepStagerNetV4: Single-Channel EEG Sleep Staging

Breaking the κ = 0.57 Barrier in Automated Polysomnography via Four Synergistic SOTA Innovations

in plain wordsScoring a night of sleep by hand takes a technician hours. SleepStagerNetV4 reads sleep stages from a single EEG sensor and attacks the four weaknesses that hold baseline models at κ ≈ 0.57, aiming for the agreement of human experts (κ ≥ 0.81).

κ ≥ 0.81: Target Cohen's Kappa
T = 21 s: Temporal Context Window
4,999 Epochs: Labelled Clinical PSG
NVIDIA H100: Hardware Benchmark

Manual polysomnography (PSG) scoring requires hours of painstaking technician labor, while expert inter-rater agreement sits at Cohen's κ ≈ 0.76–0.83. Automated single-channel EEG analysis drastically minimizes patient sensor discomfort but is crippled by the N1/REM latent collapse (both stages feature low-amplitude theta dominance without chin EMG).

SleepStagerNetV4 targets four systematic bottlenecks of baseline models (κ ≈ 0.57) to achieve clinical inter-rater parity (target κ ≥ 0.81) across 4,999 labelled epochs and 1,000 test epochs under the 5-class AASM taxonomy (Wake, N1, N2, N3, REM).

What’s new in it

Differentiable STFT & Cross-Attention STFT + Fusion
Fixes time-domain blindness to sleep spindles (12–15 Hz) & delta rhythms (0.5–4 Hz). On-the-fly STFT branch queried by multi-scale temporal convolutions (Q = z_time, K = V = z_freq; expected Δκ: +0.05–0.08).
Supervised Contrastive Loss SupCon
Resolves N1/REM latent collapse by pulling same-class pairs together and cross-class theta-dominant pairs apart on the latent hypersphere: L = L_Focal + 0.10·L_SupCon (expected Δκ: +0.03–0.05).
Extended 21-Epoch Context T=21 Context
Expands temporal perception from 7s to 21 consecutive 1s epochs (~70% of an AASM 30-s clinical scoring epoch), captured via 8-head MHSA and 2-layer BiLSTM (expected Δκ: +0.04–0.07).
Adaptive Per-Class Thresholds FlexMatch
Dynamic confidence thresholding: τ_c(t) = τ_base·(σ_c(t) / max σ_c'(t)). Prevents minority stages (Wake 10.2%, N1 13.9%) from pseudo-label starvation (expected Δκ: +0.03–0.06).

The guarantees, the code and the specs

5-Class AASM Taxonomy: Wake (Alpha 8–13 Hz, 10.2%), N1 (Theta 4–8 Hz, 13.9%), N2 (Spindles 12–15 Hz, 40.0%), N3 (Delta 0.5–4 Hz, 20.8%), REM (Sawtooth waves, 15.0%). Imbalance ratio: 3.9×.
Zero Data Leakage Protocol: StandardScaler fitted strictly per training fold; SupCon operates exclusively on labelled mini-batches; FlexMatch thresholds adapt solely to unlabelled confidence.
5-Pass Noise-Augmented TTA: Inference employs 1 clean pass + 4 noise-augmented passes (σ = 5%) with zero label information leakage.

SleepStagerNetV4 System Architecture Pipeline:

Input: Context window T = 21 consecutive EEG epochs (178 Hz, 1s each)
  ├─ ResidualMultiScaleEncoder (kernels 3 / 7 / 15)  → z_time  (256-d)
  ├─ SpectrogramBranch (torch.stft n_fft=32, hop=4)   → z_freq  (256-d)
  └─ CrossAttention (Q = z_time, K = V = z_freq)      → z_fused (256-d)
        ↓ Sequence of 21 × 256-dim embeddings
  Multi-Head Self-Attention (8 heads, d=256, T=21)
        ↓
  Bidirectional LSTM (2 layers, 128 hidden/direction)
        ↓ Mean-pool over time
  FC(256→128) → LayerNorm → GELU → Dropout(0.25) → FC(128→5) logits
  Auxiliary: Centre-epoch embedding → SupCon Loss (τ=0.07, weight=0.10)

Stage	Count	Proportion	Dominant Rhythm	Clinical Neurophysiological Markers
Wake (0)	511	10.2%	Alpha (8–13 Hz)	Low-amplitude mixed-frequency; beta in active wakefulness
N1 (1)	697	13.9%	Theta (4–8 Hz)	Transitional; alpha loss; vertex sharp waves
N2 (2)	2,000	40.0%	Sigma/Spindles (12–15 Hz)	Sleep spindles (0.5–2s bursts); K-complexes
N3 (3)	1,040	20.8%	Delta (0.5–4 Hz)	High-amplitude slow-wave delta; deepest NREM sleep
REM (4)	751	15.0%	Mixed / Sawtooth	Low-amplitude theta-dominant; sawtooth waves

ch. 05 — education

Education

Five years at Algeria’s national school of artificial intelligence, written up like a transcript.

State Engineering Degree in Artificial Intelligence & Data Science

National Higher School of Artificial Intelligence (ENSIA) · Algiers · 5-year integrated programme (MSc equivalent)

2021 – 2027

Higher Engineering Specialization in AI & Data Science

2024 – 2027 · final year, project in progress

Advanced engineering specialization focused on statistical machine learning, low-resource Arabic NLP, and autonomous intelligent systems. Completing specialized coursework while undertaking research-driven engineering projects leading to the State Engineering Degree.

Core courses
Deep Learning	Neural architectures, transformer models, representation learning, optimization, and RAG retrieval pipelines.
Natural Language Processing	Low-resource dialect modeling, Arabic NLP semantics, multilingual tokenization, and sequence modeling.
Computer Vision	Visual feature representations, convolutional networks, object detection, and spatial understanding.
Reinforcement Learning	Markov decision processes, policy optimization, reward function engineering, and agentic systems.

Pre-Engineering Preparatory Cycle

2021 – 2024 · completed

Three years of rigorous national engineering foundation in higher mathematics, theoretical computer science, and algorithmic problem-solving.

Foundation subjects
Higher Mathematics	Multivariable Calculus, Linear Algebra, Probability Theory & Applied Statistics
Computer Science	Data Structures, Graph Theory, Algorithm Complexity & Optimization
Systems & Engineering	Computer Architecture, Systems Programming in C/C++, and Software Engineering

Admitted through the national competitive entrance ranking.

ch. 06 — conferences

Conferences

Academic conferences and scientific research presentations attended

1ST EDITION • 2025

AAIET'25

Applied AI & Emerging Tech

Ziane Achour University of Djelfa

International Conference on Applied AI and Emerging Technologies

Participated in peer-reviewed scientific paper presentations and keynote lectures exploring practical artificial intelligence applications, emerging neural architectures, and data engineering innovations across Algerian and international research laboratories.

the tracks

Applied Artificial Intelligence — Practical ML deployment, predictive modeling, and domain adaptation in low-resource environments.
Emerging Neural Architectures — Deep learning model advancements, representation learning, and modern computer vision/NLP architectures.
Research & Industry Exchange — Connecting academic research findings with industry datathons and scalable software engineering.

ch. 07 — certificates

Certificates & Attestations

Professional credentials, certifications, and event participations arranged by impact

Industry-recognized certifications and professional credentials

Hackathons, datathons, competitions, workshops, and major events attended

Completed online courses and self-directed learning accomplishments

Tokyo, Japan

Advanced AI Track Matsuo-Iwasawa Lab

The University of Tokyo (東京大学松尾・岩澤研究室)

Intensive training by the renowned Matsuo-Iwasawa Lab at The University of Tokyo. Deep curriculum covering state-of-the-art Generative AI, Large Language Models (LLMs), multimodal foundation architectures, prompt optimization, and autonomous multi-agent systems with rigorous implementations.

Core Focus: Generative deep learning, LLM reasoning pipelines, multimodal transformers, and production AI agent workflows.

Generative AI LLMs PyTorch AI Agents

Course Completed Lab Portal ↗

Google Cloud · MENA

Cloud & AI Engineering Launched Sep 2026

Lead Instructor: Abdel Sghiouar (Senior Cloud Developer Advocate @ Google Cloud)

5-week practical technical training track designed for developers and AI engineers across the MENA region. 100% hands-on coding with zero slide decks, delivered entirely in Arabic.

Season 1 Core Focus: • Cloud Infrastructure: Building & deploying scalable microservices on Google Cloud • Google Antigravity: Utilizing the Google Antigravity framework for cloud workflows • AI Agents: Autonomous reasoning agents built with Agent Development Kit (ADK) on Cloud Run

Google Cloud Google Antigravity AI Agents (ADK) Cloud Run

100% Hands-on Coding Program Portal ↗

ch. 08 — languages

Languages

Three languages, shaded in on the usual A1–C2 scale.

Arabic العربية

MSA & Algerian Darija

A1A2B1B2C1C2

native

My mother tongue — at home, and with the Algerian teams and users I build for.

English

Professional & academic

A1A2B1B2C1C2

C1–C2

What I read research in, write documentation and papers in, and work in day to day.

French Français

Professional working

A1A2B1B2C1C2

B2–C1

The language of my engineering studies at ENSIA, and of most coursework and case competitions.

appendix — tech terms

The words on this page, in plain English

Some of what’s above is jargon. Here’s what it means — no background needed. Anywhere on the page, a dotted word opens its definition right where you are.

Agentic AI n.

AI that doesn’t just answer — it takes steps. It looks things up, uses tools, checks its own work, then acts. Think of an assistant who goes and does the errand rather than describing how.

Texa’s trip-planning agents; the healthcare system that asks clinicians questions.

Competitive programming n.

Solving algorithm puzzles against the clock, scored automatically. A sport for programmers — and the best training there is for writing fast, correct code under pressure.

Coaching students for the Algerian Olympiad in Informatics.

Deepfake n.

A fake image, video or voice made by AI to look and sound real. Detecting them is a research problem in its own right.

My research collaboration with a professor at Montreal University.

EEG n.

Electroencephalography: a recording of the brain’s electrical activity, picked up by sensors on the scalp. Sleep doctors read it to tell which stage of sleep someone is in.

SleepStagerNetV4 does that reading automatically, from a single sensor.

Faithfulness n.

How closely an AI’s answer sticks to its sources instead of inventing things, scored from 0 to 1. A score of 1 would mean every claim is backed by a document.

SPIRAL-RAG scores 0.856 on 330 official legal questions.

Full-stack adj.

Building both halves of an app: the part you see and click (the frontend) and the part behind it — server, database, security (the backend).

Most of the client work and internships here.

Hallucination n.

When an AI states something confidently that simply isn’t true — a road that doesn’t exist, a citation nobody wrote. Most systems on this page are built specifically to prevent it.

The reason Texa is tied to official Ministry data.

LLM n.

Large language model: the kind of AI behind ChatGPT, trained on vast amounts of text to read and write language. Long-context means it can take in very long documents at once — a whole contract, not a page.

Gemini Pro and Llama 3.3 in SPIRAL-RAG; my current research.

Machine learning n.

Teaching a computer by example instead of by rules: show it many examples and it learns the pattern itself — the way you learnt to recognise a cat without a definition of one.

The foundation under everything on this page.

Multi-agent system n.

Several AI agents, each with one job, working together — one plans, one checks facts, one fetches data — like a small team rather than one person doing everything.

Texa, the national tourism platform.

NL-to-SQL n.

Turning a question in plain language (“how many subscribers left last month?”) into a database query, so people who don’t write code can still ask their data questions.

The data-science internship at Djezzy.

Preprint n.

A research paper shared publicly before, or while, it goes through formal peer review — so others can read and build on it sooner.

QPNN-v2 and MoQAdam.

RAG n.

Retrieval-augmented generation. Before answering, the AI looks up the relevant documents and answers from them — an open-book exam instead of answering from memory. It’s the main defence against hallucination.

MedRAG Ultra, SPIRAL-RAG, and my research at FEMTO-ST.

Reinforcement learning n.

Learning by trial and reward, the way you might train a dog: whatever leads to a good result gets reinforced. RL for RAG uses it to teach a system when to ask a clarifying question instead of guessing.

MedRAG Ultra, RAMY, and current research.

State estimation n.

Working out where a robot is and what is around it from noisy, partial sensor readings — like knowing where you are in a dark room from what you can touch.

RAMY, the robotics agent.

Vector database n.

A database that finds things by meaning rather than exact words, so a search for “car” also finds “automobile”. It’s what a RAG system searches.

MedRAG Ultra’s retrieval.

still unclear about something? Ask me — I’m happy to explain

ch. 09 — contact

Let's Build Something Together

Freelance work, a role on your team, or research — write a few lines below and I’ll answer myself.

Thanks for reading this far. Samir

AI engineer & full‑stack developer.

What I'm building right now

VOXORA — BeLogic AI

Industrial agentic system for healthcare

Deepfake AI detection

TEXA — Ministry of Tourism

Filand — decision support

Flagship Systems & Deployed Platforms

Built at work

VOXORA // Voice & Chat AI Agents for European Businesses

how it works One conversation, through the platform

Built for institutions & research

Texa // National Tourism Intelligence

MedRAG Ultra // Clinical Decision Support

RAMY // Intelligent Agent & Robotics

SPIRAL-RAG // Regulatory Archive Intelligence

Built for clients, freelance

SkinDZ // Cash-on-Delivery Beauty Commerce

Doctor Darna // Home Healthcare Platform

El Basma // Storefront, Back-Office & AI Landing Pages

v1 → v2 The whole shop, rebuilt

El Basma App // The Team’s iOS & Android Back Office

Dahaoui Boutique // Black-and-Gold Storefront & Back-Office

before → after From a template to a boutique

Akkida // Order Confirmation & Delivery Operations

Where I’ve worked

AI & Agentic Engineer

Data Scientist Intern

Full-Stack Software Engineer Intern

Network & Systems Engineer Intern

Frontend Developer Intern

Client work, on the side

Dahaoui Boutique

Self-Employed

Teaching & coaching

Competitive Programming Coach

Algorithms Teaching Assistant

IT Manager & Technical Lead

Research

QPNN-v2: QuantumPath Neural Networks

What’s new in it

MoQAdam: Moment-Normalized Quantization Adam

What’s new in it

SleepStagerNetV4: Single-Channel EEG Sleep Staging

What’s new in it

Education

State Engineering Degree in Artificial Intelligence & Data Science

Higher Engineering Specialization in AI & Data Science

Pre-Engineering Preparatory Cycle

Conferences

International Conference on Applied AI and Emerging Technologies

Certificates & Attestations

Industry-recognized certifications and professional credentials

Hackathons, datathons, competitions, workshops, and major events attended

Languages

Arabic العربية

English

French Français

The words on this page, in plain English

Agentic AI n.

Competitive programming n.

Deepfake n.

EEG n.

Faithfulness n.

Full-stack adj.

Hallucination n.

LLM n.

Machine learning n.

Multi-agent system n.

NL-to-SQL n.

Preprint n.

RAG n.

Reinforcement learning n.

State estimation n.

Vector database n.

Let's Build Something Together