Achraf HafsaouiDATA & AI
Data & AI

Thoughtful software.Intelligent systems.Useful outcomes.

I’m Achraf. I work across software engineering and machine learning, turning complex problems into tools people can use.

Explore my work GitHub
01 / MY ROADMAPSTUDY + PRACTICE

Building on each chapter.

Computer science, artificial intelligence, and applied work.

  1. OCT 2021 — JUN 2025

    B.Sc. Computer Science

    TU Darmstadt

    Software engineering foundations
  2. APR 2025 — PRESENT

    M.Sc. AI & Machine Learning

    TU Darmstadt

    Deepening the theory
  3. JUL 2026 — PRESENT

    Data & AI Working Student

    Abbott

    Putting it into practice alongside my master’s
02 / Where I’m working

AI, in practice.

JUL 2026 — PRESENT

Abbott

Diagnostics · Technical Operations

Data & AI Working Student

Wiesbaden, Germany

Building a RAG application that lets colleagues ask questions and find the relevant passages across thousands of documents.

Collaborating with the cloud team to deploy the application on Azure Kubernetes Service, with database management, authentication, and access control.

Working directly with subject-matter experts to understand what they need to find and how they would use the application.

Retrieval-augmented generation

Making unstructured
knowledge usable.

KNOWLEDGE BASEThousands of documentsUnstructured source material
RETRIEVALThree complementary channelsCollect relevant candidate passages
RERANKINGPrioritize the relevant evidenceSource context for AI-assisted answers
Python Azure Kubernetes Service (AKS) CloudNativePG (PostgreSQL) Docker GitHub Actions Microsoft Entra ID
03 / Personal & Academic Projects

Different problems.
The same curiosity.

01 / CI/CD · Conceptual flow
// A failed build is the starting point.
on push → analyze(build_logs)
retrieve(relevant_context)
generate(suggested_fix)
return pull_request_comment
Personal project03/2026 — Present

AI-Powered CI/CD Debugger

A distributed agent that investigates failed builds and brings suggested fixes back to the pull request.

PythonGoPostgreSQL (pgvector)RabbitMQDockerStreamlitLLMsGitHub Actions
GitHub ↗
02 / Code generation · Proposed architecture
Individual seminar04/2026 — 09/2026

Constrained Decoding

Exploring a fast/slow approach to code generation: inexpensive checks supported by deeper, asynchronous syntax analysis.

LLMsFinite-state machinesAST analysis
Read paper PDF ↗
03 / Multimodal AI · Illustrative task
“When does the person open the door?”
VIDEO STARTMATCHING INTERVALEND
Multimodal AI project lab10/2025 — 03/2026

Video Temporal Grounding

Investigating how multimodal models locate events in video through prompt design, benchmarking, and output processing.

“Across Charades-STA, ActivityNet Captions, and QVHighlights, we achieve new state-of-the-art performance among all Training-Free approaches, improving R1@0.5 by +7.5%, +9.2% and +10.9%, respectively—closing the gap to fine-tuned methods.”

Research Conclusion
PythonPyTorchHugging FaceLLMs (Qwen, GPT)Prompt EngineeringPost-processing Techniques
Read paper PDF ↗
04 / Reinforcement learning · Agent loop
PIXELSOBJECTS
State
→
PQN agent
→
Action
↘ ENVIRONMENT → REWARD + NEXT STATE ↗
Reinforcement learning10/2025 — 03/2026

Atari RL: Double Dunk

Trained and benchmarked reinforcement learning agents on Double Dunk using pixel-based and object-based representations.

PythonJAXReinforcement LearningWeights & Biases (W&B)PQN Agents
GitHub ↗
05 / Fleet monitoring · Application architecture
Team project10/2024 — 03/2025

Bus Monitoring Platform

A real-time platform for monitoring and managing electric bus fleets, developed with a team as an MVP for an early-stage startup.

“Overall, I found your work to be structured, reliable, and technically sound. You quickly adapted to new topics and contributed effectively to the team in solving complex tasks.”

Dr. Ing. Benjamin Blat Belmonte — Founder, Fenexity
TypeScriptReactTailwind CSSDjangoPostgreSQLViteREST-APIs (Axios)Postman (Tests)Jira (Scrum)
Frontend ↗Backend ↗
06 / Collaboration · Shared state
Team project10/2023 — 03/2024

Cross-Platform Collaborative App

A cross-platform application where groups collaborate through real-time group chats and shared whiteboards, keeping conversations and visual ideas in one workspace.

JavaScriptReact NativeNode.jsMongoose (MongoDB)REST-APIs (Axios)WebSockets (Socket.IO)Jest (Tests)CSSTrello (Scrum)
GitHub ↗
04 / A little background

Engineer by training.
Curious by nature.

I like understanding how things work and making them work together.

I’m a Data & AI Working Student based in Germany, pursuing a master’s in Artificial Intelligence and Machine Learning at TU Darmstadt .

My interests run from retrieval and code generation to the software around them: interfaces, APIs, databases, and the details that make a system useful. Alongside my studies, I build data and AI tools at Abbott .

AI & research

Python, PyTorch, Hugging Face, LLMs (Qwen, GPT), JAX, Reinforcement Learning, PQN Agents, Prompt Engineering, Post-processing, Multimodal Models, Constrained Decoding, Finite-state Machines, AST Analysis

Retrieval & RAG

RAG, BM25, Dense Embeddings, Hybrid Retrieval, Reranking, Information Retrieval, Document Processing

Backend & data

Go, Node.js, Django, PostgreSQL, pgvector, Mongoose (MongoDB), REST APIs, WebSockets (Socket.IO)

Frontend & mobile

TypeScript, JavaScript, React, React Native, Tailwind CSS, HTML, CSS, Vite, Streamlit, Axios

DevOps & testing

Docker, Kubernetes, Git, GitHub Actions (CI/CD), RabbitMQ, Weights & Biases (W&B), Jest (Tests), Postman (Tests)

Collaboration

Agile (Scrum, Jira, Trello), Stakeholder Communication, Knowledge Sharing

Project notes
Constrained Decoding · Architecture

FSM masking, tripwire detection, asynchronous AST construction, and rollback. Scroll horizontally on smaller screens to inspect the full diagram.
Bus Monitoring Platform · Architecture

Role-based access, application services, data storage, and electricity-price integration. Scroll horizontally on smaller screens.
Cross-Platform Collaborative App · Architecture

Group chats and shared whiteboards synchronize through Socket.IO and Node.js. Mongoose connects the backend to MongoDB for data storage.