Elyas El Khaldi
MSc Data Science student at EPFL (2025–2027). Working on how language models are fine-tuned, retrieval for code repair, and agents that can be evaluated properly.
Now
I'm a semester-project student at the EPFL Data Science Lab, studying whether behaviours attributed to fine-tuned LLMs belong to the models or to the adaptation method. I also write embedded C firmware for a CubeSat power unit with the EPFL Spacecraft Team, and I've been a teaching assistant at EPFL since 2023.
In summer 2026 I built and evaluated multi-agent workflows as a Generative AI Consultant Intern at Simon-Kucher & Partners.
Education
2025 – 2027EPFL, MSc in Data ScienceGrade 5.74/6
2022 – 2025EPFL, BSc in Communication SystemsGrade 5.78/6, rank 4/196
2024 – 2025Nanyang Technological University, SingaporeExchange year
Recognition
EPFL Master Excellence Fellowship, awarded to 33 of 1,000+ incoming Master's students.
Contact
Work
Experience
2026 – PresentSemester Project, EPFL Data Science Lab (DLab)Research
- Investigating how LoRA versus full-parameter fine-tuning affects learned behaviours and persona-direction mechanisms in LLMs, and whether reported effects are intrinsic to the models or artifacts of the adaptation method.
- Built a reproducible PyTorch, Transformers and PEFT pipeline for a ~90-run Qwen3-8B study, varying fine-tuning configurations, learning rates and random seeds.
Jun – Aug 2026Generative AI Consultant Intern, Simon-Kucher & PartnersInternship
- Built and evaluated LangGraph multi-agent workflows for consulting use cases, turning client requirements into reproducible AI systems.
- Developed an evaluation harness with a golden dataset, task-success metrics, end-to-end tracing and human-in-the-loop review to compare agent architectures.
2025 – PresentSoftware Engineer, EPFL Spacecraft TeamFirmware
- Developed embedded C firmware for a CubeSat Electrical Power Unit.
- Debugged hardware/software interactions and contributed to a shared codebase under reliability and resource constraints.
2023 – PresentTeaching Assistant, EPFLTeaching
- Supported 5 courses for 5 professors.
- Explained technical topics to 100+ students, led tutorials and assessed coursework.
Projects
LLMsPreAct: LLM code repair and retrievalResearch
- Repository-scale code-repair pipeline combining retrieval, a LoRA-fine-tuned Qwen2.5-0.5B reranker and AST-based source analysis to find edit-relevant files.
- Trained on SWE-smith and evaluated on SWE-bench Verified: file-localisation Recall@3 rose from 43.3% to 58.3% and Recall@5 from 60.0% to 71.7% over BM25.
Web3Stars: on-chain ticketing mini-appCode ↗
- End-to-end on-chain ticketing app shipped as a mini-app inside Coinbase's Base social app.
- Won Start Hack Tour × St. Gallen and was a finalist in the Base Builder Track.
DataYouNiverse: large-scale network analysisCode ↗
- Built a YouTube audience-overlap graph of 19,129 channels and 572,732 edges.
- Louvain community detection found 52 communities with a modularity of 0.655.
AgentsSIGMA: self-improving greenhouse Mars agentSimulation
- A Mars environment simulation and an autonomous AI agent that manages a greenhouse and its resources for human survival.
Miscellaneous
Outside work
SportFootball, Brazilian jiu-jitsu
FilmCinema
LanguageLinguistics
Languages
FluentFrench, Arabic, English
WorkingSpanish, Italian
Hackathons
2026Start Hack Tour × St. Gallen, winner (Stars)
2026Base Builder Track, finalist (Stars)
2025Start Hack 2025, participant
Toolbox
LanguagesPython, C, C++, Java, Scala, SQL, TypeScript
MLPyTorch, Hugging Face Transformers/PEFT, LangGraph, RAG, LoRA and full fine-tuning
SystemsLinux, Git, Docker, React, embedded systems, low-level debugging