AI engineering student · ESPRITTunis, Tunisia / 2026

Melek Moalla

I build AI.
Then I test
its limits.

Hands become interfaces. Agents learn to cheat. Models learn to remember. I build these systems—and measure what actually works.

Open to AI / ML internships

Melek Moalla
Melek MoallaStudent. Builder. Always testing.
Computer vision / Reinforcement learning / AI agentsGitHub ↗LinkedIn ↗CV ↗
01 / Selected work

Built to be seen.
Backed by evidence.

Start with the demo. Open a case study for the decisions, experiments, and limits behind it.

Web & more ↘

Showing 8 projects

01 / Reinforcement learning2026

Recorded PPO policy · side-by-side with the intended solution

Reward Goblin

A physics playground where real PPO agents optimise the reward you wrote instead of the task you meant.

53.5 vs 3.9reward earned by the cheating policy vs the correct solution · Edge Goblin v1, seed 1
Explore case study →
What to keep in mind

Exploit detectors are hand-tuned heuristics: diagnostics, not proofs. Expect false negatives on rewards they were not tuned for.

02 / Computer vision2026

Real gesture engine · synthetic hand input

Vector

A webcam becomes a spatial input device: pinch, drag and throw real Windows windows with your hands.

1 in 11.5 minaccidental commands under adversarial synthetic motion; zero clicks, drags or throws
Explore case study →
What to keep in mind

The demo is the real HUD painter and gesture engine driven by a synthetic hand. A live screen recording is still to be made.

03 / Agents · Systems2026
Hydra · manager agent orchestrating a fleet over MCPillustrative fleet · real tool names
Hydra orchestration diagramA human uses the Hydra desktop app. A manager agent runs an MCP server inside it and calls tools that act on up to ten agent terminals. Selected tool: hydra_broadcast. Send the same prompt to every agent in one project.Youchat view · grid viewManager agentMCP server · HTTP/SSEhydra_broadcast()Headless runsJSONL logs · resumeclaudepty · project apiclaudepty · project apicodexpty · project apiclaudepty · project webcodexpty · project webnew agentspawned on demandproject apiproject web

hydra_broadcast — Send the same prompt to every agent in one project.

Interactive architecture · illustrative fleet, real MCP tools

Hydra

A desktop app that runs Claude Code and OpenAI Codex agents in parallel, with a manager agent that coordinates the others over MCP.

7 MCP toolslet one manager agent spawn, prompt, read and restart up to 10 concurrent agents
Explore case study →
What to keep in mind

Usage and cost figures are parsed from CLI output: signals, not billing data.

04 / Reinforcement learning · Computer vision2026

Learned policy · synthetic teacher demonstrations

GhostHands

Demonstrate a pick-and-place with your hand, change the scene, and test what a learned policy actually generalises.

46% → 97%held-out success of the same MLP when raw poses become relative geometry · 100 paired layouts
Explore case study →
What to keep in mind

The included checkpoints use synthetic teacher demonstrations, not human webcam data. Human recording is implemented but unvalidated with a participant.

05 / Computer vision2026
A red mug lifted out of the scene by a pinch, with its original spot filled in

Real pipeline · synthetic desk fixture

Telekinesis

Touch a real object in the webcam image, pinch, and lift its appearance out of the video while the background is reconstructed behind it.

0.3–1.1 sEdgeSAM encode on a laptop CPU, run once per selection; tracking is per-frame
Explore case study →
What to keep in mind

Segmentation varies with contrast, clutter, thin, transparent or reflective objects.

06 / Reinforcement learning2026
Heatmap of the GRU hidden state across 34 decisions in the intact control episode

Recorded hidden state · memory-intact episode

Dungeon With Amnesia

A PPO laboratory that tests whether agents remember what they can no longer see, by deleting their memory mid-episode.

100 → 43escapes out of 100 when the GRU's hidden state is reset mid-episode
Explore case study →
What to keep in mind

The strong result concerns retaining a one-time cue in a controlled detour, not learned memory of a door's location.

07 / Systems · Agents2026
Live dashboard with KPI tiles, incidents by hour and anomaly trend

Working dashboard · synthetic payment events

Payment Observability

A real-time dashboard that replays synthetic payment events through an event bus, a risk engine and a guarded conversational analyst.

100% syntheticreplayed through an event bus, a risk engine and a Redis cache into a live dashboard; no real records
Explore case study →
What to keep in mind

The public version runs on generated demo data. It demonstrates patterns, not production scale.

08 / Applied ML2025–2026
Visual flow from naive DNA to protein structure to host-optimised DNA

DNA design pipeline · simulated expression target

Protein Plant Synthesis

A leakage-controlled surrogate ML pipeline that designs DNA for therapeutic proteins across four plant hosts, and explains its own predictions.

4 hostsArabidopsis, rice, maize and date palm, compared under protein-grouped evaluation
Explore case study →
What to keep in mind

The expression target is a deterministic simulated proxy built from codon, host and structure signals. It is explicitly not wet-lab measured.

02 / Behind the work

Curiosity first.
Rigor follows.

I'm an AI engineering student at ESPRIT in Tunis. I like problems where the obvious metric lies.

Agents that game rewards. Gestures that trigger by accident. Models that only work on the layout they trained on. Building the system is half the work; finding out where it fails is the other half.

I work with coding agents as collaborators and review what they produce. That experience led me to build Hydra.

ESPRIT / 2024–25First in my class.

Annual class ranking. AI engineering degree expected in 2028.

Learning / NVIDIAML. Deep learning. Generative AI.

Certificates supporting the work you see here.

Experience / Three internshipsFrom interfaces to AI systems.

MS Solutions Group · Telnet · Shamash IT

Education & experience
  1. Engineering degree, Computer Science, Artificial Intelligence trackESPRIT · expected
  2. Reward Goblin, Dungeon With Amnesia, Vector, Telekinesis, GhostHands, HydraAn intense month of building, all public on GitHub
  3. Software & AI Intern, MS Solutions GroupReal-time transaction monitoring with FastAPI, WebSockets and Redis; an LLM chatbot over dashboard data
  4. AI specialisation year · NVIDIA certificates in Machine Learning, Deep Learning and Generative AIPlant DNA deep-learning project selected for the Bal des Projets showcase
  5. Ranked 1st in the annual class rankingLogiXpress wins at the Bal des Projets; WaveWorx selected
  6. Software Intern, TelnetBuilt and deployed a Docker-based live metrics dashboard from supervisor specifications
  7. Ranked 2nd in the annual class rankingLinux game and Arduino controller selected for the Bal des Projets
  8. Web Development Intern, Shamash ITUser-facing interfaces in HTML, CSS and JavaScript
  9. Baccalaureate in MathematicsLycée El Menzah 9
03 / The wider lab

Still exploring.

Earlier work, team projects, and what I'm building next.

Showing 6 lab projects

2026 / In progress · not yet public

ATLAS

Phone video in, metric 3D digital twin out. Today it recovers camera poses with COLMAP, estimates metric scale without a depth sensor (a learned geometry prior plus a bootstrap confidence interval) and fuses a TSDF mesh, on CPU, measured against ground truth. Splats, rooms and floor plans are still planned.

Tools used

Python · COLMAP · Open3D · OpenCV

In progress
2025–2026 / Multi-module integration project

Plant DNA & crop intelligence platform

A Flutter field app and a React research dashboard behind one FastAPI gateway serving a wheat-disease CNN, a pest image classifier, an insect-sound classifier and a FAISS-backed retrieval workflow for DNA search.

Tools used

TensorFlow · PyTorch · FastAPI · Flutter · React · FAISS · LangChain

Source
2024–2025 / University project · Bal des Projets winner

LogiXpress

A PHP/MySQL MVC delivery platform with route views, packaging orders and claims, plus experimental Python sidecars for delivery forecasting and biometric login.

Tools used

PHP · MySQL · Python · scikit-learn · Leaflet · Firebase

Source
2025 / Team project

Festy Event

Event booking for Tunisia in Django: QR-code tickets, a Leaflet map of venues by governorate, availability calendars, PDF invoices with VAT and admin analytics.

Tools used

Django · Python · Leaflet · Chart.js · SQLite

Source
2024–2025 / University project · Bal des Projets selection

WaveWorx

A C++/Qt operations desktop app on Oracle: inventory and business records with search, stock alerts, charts and PDF reports.

Tools used

C++ · Qt · Oracle

University project
2023–2024 / University project · Bal des Projets selection

Linux game + Arduino controller

A game written for Linux, played with a gamepad I built on an Arduino board.

Tools used

C · Arduino

University project
04 / Let's build something

Have a hard
problem?

I'm looking for AI / ML internships and teams where I can build, test, and keep learning.