Ahmed RebaiGet in touch
Ahmed Rebai
rebaiahmed244@gmail.comWhatsAppLinkedInGitHub

Ahmed Rebai

Data EngineerAI · Cloud · Data Systems

I build data and AI systems the way good pipelines are built: every output traceable, every change tested, every deploy automated.

Explore my work View GitHub
Ahmed Rebai

Welcome to my portfolio. I'm Ahmed Rebai, an AI and data engineer. Let's work together and build useful software products.

Status
Open to work
Based in
Sfax, Tunisia
Studying
Data Engineering, FSS · 2028
Certified
DP-700 · DP-600 · DP-900

01About

Pipelines, platforms and grounded AI.

I'm a second-year Data Engineering student at the Faculty of Sciences of Sfax, working where data platforms, cloud infrastructure and AI meet.

In 2026 I shipped RAG, multi-agent and edge-AI systems across three internships: retrieval over knowledge graphs, a hand-built ReACT loop backed by 421 tests, and a TinyML model running on a smartwatch and a microcontroller.

What interests me is the part between the model and the user: lineage, validation, observability. I want systems where a claim can be traced back to its source and a failure is caught by code, not by a reviewer. I hold three Microsoft data certifications, including Fabric Data Engineer Associate, and I'm heading toward data platform and AI infrastructure engineering.

02Experience

Where I've shipped.

Four engineering roles across AI systems, edge ML and data backends, three of them in 2026.

  1. Aug 2026

    Hybrid

    Software & Edge AI EngineerCompiTechnology

    End-to-end delivery from on-device ML and firmware to a containerized data backend and live dashboard.

    • Integrated a TinyML model (Edge Impulse, TFLite Micro) into a Wear OS app via C++/NDK/JNI and ported it to an Arduino nRF52, optimizing inference to ~2.5 s on constrained hardware.
    • Designed a custom BLE/iBeacon protocol and state machine; wrote low-power firmware for ESP32-C6/XIAO (light sleep) and an Arduino variant.
    • Built a Laravel 12 REST API on SQLite (tracking, violations, sessions, inactivity logic) with PHPUnit tests, containerized with Docker.
    • Shipped a real-time React 18 + Inertia dashboard (live stats, analytics, theming) fed by polling and WebSockets.

    TakeawayFull-cycle delivery from embedded edge to data backend, and agreeing one protocol across three codebases.

    • Edge Impulse
    • TFLite Micro
    • C++ / NDK / JNI
    • Wear OS
    • Arduino
    • ESP32
    • BLE / iBeacon
    • Laravel 12
    • PHP
    • SQLite
    • React
    • Inertia
    • Docker

  2. Jun – Jul 2026

    Remote

    AI Automation Engineering InternNovagate Solutions

    The automation and retrieval core of a multi-service, multi-tenant platform.

    • Built the automation and RAG core of a multi-service platform (Next.js, FastAPI, Express, PostgreSQL), grounding LLM output in a knowledge-graph memory layer and real-time web search.
    • Implemented provider-agnostic LLM routing with retry-with-backoff, plus auth, billing and DNS-automation integrations across a shared-database, multi-tenant architecture.
    • Closed real security gaps before delivery: broken workspace authorization and webhook forgery on an unsigned third-party callback.
    • RAG
    • Knowledge graphs
    • Next.js
    • FastAPI
    • Express
    • PostgreSQL

  3. Jul 2026

    On-site

    AI Systems Engineering InternDESLAB

    A multi-agent system that writes evidence-grounded reports, with the test and CI discipline to back it.

    • Contributed to a multi-agent AI system (Flask, Vue 3, PostgreSQL, Redis/RQ) that writes evidence-grounded reports by orchestrating LLM agents through a hand-built ReACT loop, without LangChain.
    • Built retrieval over a temporal knowledge graph and live web search so every generated claim traces to a real source, plus a mathematically computed consensus score.
    • Backed it with a 421-test suite (real Postgres/Redis integration tests) and a containerized GitLab CI/CD pipeline: lint → build → test → deploy.
    • Multi-agent systems
    • ReACT loops
    • Knowledge graphs
    • Python
    • Flask
    • Vue 3
    • PostgreSQL
    • Redis
    • Docker
    • GitLab CI/CD

  4. Aug 2025

    Hybrid

    Software Engineering Intern · ASP.NET CoreEngineering & Consulting “E&C”

    A blog and CMS platform on the Microsoft web stack.

    • Developed Nova Write, a blog/CMS platform on ASP.NET Core MVC 8 with EF Core and SQL Server: CRUD, a relational schema and role-based access control.
    • ASP.NET Core
    • C#
    • SQL Server

03Work

Systems that show their work.

Selected open-source projects, strongest first, each with the architecture behind it.

Fig. 02 — Lineage engineThe systems on this page as one graph: raw sources, the step that transforms them, and what gets served. Pick an output to trace where it came from.
lineage("postmortem.md") = {logs, alerts, deploys, commits}

built in Postmortem Autopilot

P/01Flagship · LLM reliability

Postmortem Autopilot

Rebuilds an incident timeline from logs, alerts, deploys and commits, then writes a postmortem where every factual claim is mechanically traceable to a timestamped source event, or it doesn't ship.

Problem
An LLM-written postmortem reads well but can state a cause nobody can verify. In an incident review, a hallucinated fact is a correctness bug, not a style problem.
Approach
Fence the model between two deterministic planes. Code collects evidence into a temporal knowledge graph and proposes candidate causes; the LLM only ranks and phrases them; a validator resolves every citation against Neo4j and rejects the document if a single one fails.
  • 0hallucinated citations
  • ≥95%claim coverage gate
  • 2LLM calls per run
Technical depth6 notes
  • Evidence plane: collectors normalize logs, alerts and git history into signed events; signature collapsing folds repeated log lines into one failure mode (776 lines → 1 node).
  • Candidate linking without an LLM: temporal proximity, service overlap and change-path overlap. Recall is done with code, never with a model.
  • Exactly two model calls per run in LangGraph: an Analyst that ranks competing hypotheses and names disconfirming evidence, and a Writer that cites every factual sentence.
  • Bounded repair loop: validator feedback returns to the Writer at most twice, then the run fails loudly and exits non-zero, keeping the rejected draft for inspection.
  • CI-enforced evaluation on a 10-incident golden corpus: hallucinated citations = 0, coverage ≥ 0.95, precision@1 ≥ 0.70.
  • Runs locally, in plain Docker and as a k3s Job via Helm; both images build and push from GitHub Actions on every merge to main.
  • Python
  • LangGraph
  • Neo4j
  • Knowledge graphs
  • Docker
  • k3s
  • Helm
  • GitHub Actions
  • pytest

View repository Python · updated Nov 2026

Fig. 03Three-plane architecture
  1. Evidencedeterministic · no LLM
    • Collectors
    • Normalizer
    • Graph writer
    • Candidate linker
  2. ReasoningLangGraph · 2 model calls
    • Analyst · ranks hypotheses
    • Writer · cites every claim
  3. Verificationdeterministic · Neo4j lookups
    • Resolve every [src:…]
    • Coverage ≥ 95%
    • 0 uncited claims
  4. Output
    • postmortem.md
    • evidence graph

FAIL → validator feedback → Writer · max 2×, then fail loudly

P/02Real-time data engineering

TradePulse

A fully local, production-grade streaming pipeline: synthetic OHLCV ticks flow through Redpanda, get enriched in real time with VWAP, RSI-14 and anomaly detection, and land in S3 Parquet and DynamoDB.

Problem
Seeing how a modern market-analytics stack behaves end to end (broker, stream processing, storage, monitoring) without depending on cloud infrastructure.
Approach
A seven-service Docker Compose stack that runs offline on a laptop: a GBM tick producer, a Python stream processor, storage sinks on a local AWS emulator provisioned with Terraform, and Prometheus + Grafana watching all of it.
  • 4CI/CD workflows
  • 23+unit & integration tests
  • 3Grafana dashboards
Technical depth5 notes
  • Stream processor on confluent-kafka: 60 s tumbling VWAP windows, 14-period RSI crossovers and z-score anomaly alerts, each published to its own topic.
  • Avro schemas with a v2 for schema-evolution compatibility checks against the registry.
  • Sinks buffer and flush VWAP to S3 Parquet (PyArrow) and batch alerts into DynamoDB, with configurable fault injection for resilience testing.
  • Observability: Prometheus alerting rules and three provisioned Grafana dashboards, plus a replay tool and an end-to-end throughput benchmark.
  • Documented decisions, including the migration from PyFlink to confluent-kafka (ADR-003).
  • Python
  • Kafka
  • Redpanda
  • Avro
  • AWS (S3 · DynamoDB · Kinesis)
  • Terraform
  • Docker
  • Prometheus
  • Grafana
  • GitHub Actions
  • pytest

View repository Python · updated Aug 2026

Fig. 04Streaming topology
  1. Produceevery 100 ms
    • Tick simulator · GBM
    • Kafka producer
  2. BrokerRedpanda
    • market.ticks
  3. Processstream processor
    • VWAP · 60 s windows
    • RSI-14 signals
    • z-score anomalies
  4. StoreTerraform-provisioned
    • S3 · Parquet
    • DynamoDB · alerts
  5. Observe
    • Prometheus · 15 s scrape
    • Grafana · 3 dashboards

P/03Data warehousing · OLAP

Tunisian Football Data Warehouse

Python ETL that collects, cleans and integrates Tunisian league, cup and super-cup history into a constellation schema for OLAP analytics and Power BI reporting.

Fig. 05ETL into a galaxy schema
  1. Extract
    • CSV sources
    • Normalize
  2. Transform
    • Dedupe
    • Null checks
    • Date dimension
  3. Load
    • 6 fact tables
    • 9 dimensions
  4. Serve
    • OLAP · Power BI
  • Python
  • pandas
  • ETL / ELT
  • SQL
  • SQL Server
  • Power BI

View repository Python · updated Dec 2025

P/04Mobile · Flutter

Konsol

A premium SSH client for mobile: multi-session tabs, keychain-stored Ed25519/RSA keys and remote path autocomplete globbed live on the real server over a side SSH channel, with no model and no guessing.

Fig. 06Session channels
  1. Keys
    • Ed25519 / RSA
    • Device keychain
  2. Session
    • dartssh2
    • Multi-session tabs
  3. Channels
    • Terminal · xterm
    • Side channel · remote glob
  4. UI
    • Path autocomplete
    • Contextual key bar
  • Flutter
  • Dart

View repository Dart · updated Sep 2026

More on GitHub

All repositories

04Stack

Tools, with evidence.

67 tools grouped by discipline. Select one to trace where it was actually used: roles, projects and certifications on this page.

Data engineering12

AI & LLM systems12

Cloud & infrastructure11

Backend & web11

Edge & embedded8

Languages10

Testing3

Used in a role, project or certification below

05Certifications

Certified on Microsoft Fabric.

Three Microsoft data certifications earned in 2026, from Azure data fundamentals to engineering on Microsoft Fabric.

DP-600Associate

Fabric Analytics Engineer Associate

Implementing Analytics Solutions Using Microsoft Fabric

Passed Aug 28, 2026Details

DP-900Fundamentals

Azure Data Fundamentals

Microsoft Azure Data Fundamentals

Passed Aug 1, 2026Details

  1. DP-900Aug 1, 2026
  2. DP-600Aug 28, 2026
  3. DP-700Sep 7, 2026

06GitHub

Built in the open.

Public activity, pulled from GitHub at build time and refreshed daily. No numbers typed by hand.

  • 1,563contributions in the last year
  • 98day current streak
  • 19public repositories
Contributions · last 12 months
NovDecJanFebMarAprMayJunJulAugSepOctMonWedFri
Contributions per month
MonthContributions
Nov 20250
Dec 20259
Jan 20260
Feb 20263
Mar 20260
Apr 202631
May 20260
Jun 202655
Jul 2026391
Aug 2026231
Sep 2026548
Oct 2026288
Nov 20267
Languages · share of code in public repos
  1. JavaScript44%
  2. Python22%
  3. TypeScript16%
  4. Dart9%
  5. PHP7%
  6. Other2%

Updated Nov 1, 2026 github.com/Ahmed1-Rebai

07Contact

Let's build software that shows its work.

I'm open to work in data engineering, AI systems and cloud. The fastest way to reach me is email or WhatsApp.