CNSM 2026Accepted
An Agentic AI Framework for Cloud-Native 5G Core Orchestration
L. Said Chikh, A. Mekrache, A. Cheriet, A. Ksentini
Multi-agent LLM framework for autonomous deployment and self-healing of a 5G Core on Kubernetes.
About
I work on agentic AI and large language models, and networking. What I enjoy most is getting agents to do real work on real systems, configuring, deploying and repairing. I have built agent runtimes, RAG systems, and live telemetry pipelines
I'm always up for research collaborations. If you're working on agents, LLMs or networks and want to trade ideas or build something together, drop me an email.
- Agentic AI
- LLM Agents
- RL environments
- tool use
- verification & evaluation
- benchmark & reward design
- efficient inference
News
- 2026Our paper on agentic 5G Core orchestration was accepted at CNSM 2026.
- Jun 2026Defended my MSc thesis on agentic AI for telco cloud orchestration (ENSIA × EURECOM).
- Apr 2026Joined PrevAI Health as an AI engineer (part-time).
- Dec 2025Our paper on Arabic children's story generation appeared at the Muslims in ML workshop, NeurIPS 2025.
- Nov 2025Started my MSc thesis at EURECOM on multi-agent LLM orchestration.
Publications
MusIML @ NeurIPS 2025Workshop paper
Crafting Culturally Aligned Narratives: Large Language Models for Arabic Children's Story Generation
H. Boukhalfa, S. Mani, L. Said Chikh
714-story annotated SFT dataset (κ = 0.79–0.84), LoRA fine-tuning of four open LLMs (7B–13B) and Gemini 2.0, evaluated with rule-based metrics and a panel of 5 educators (ICC = 0.85).
Preprint, 2025
Towards Semantic Understanding of Classical Arabic Poetry: Retrieval, Explanations, and Evaluation
O. F. Zouak, S. Mani, L. Said Chikh, R. A. Ouahioune
Uses LLM-generated explanations to improve embeddings for semantic retrieval of classical Arabic poetry, with evaluation by traditional metrics and LLM judges.
Experience
AI Engineer (part-time)
Apr 2026 – presentPrevAI Health
- Refactored the scaffolding of a production LLM data-science agent, then ablated prompt-engineering and prompt-compression variants to cut its context while preserving output quality.
- Built a reproducible analysis pipeline (custom parser, fully executed notebooks) over the PH-LLM fitness case-study data.
Research Intern (MSc Thesis)
Nov 2025 – Jun 2026EURECOM · Sophia Antipolis, France · Remote
- Designed a multi-agent pipeline that turns natural-language intents into Helm configurations, validates them (policy rules + Kubernetes dry-run), deploys a 5G Core (OpenAirInterface), then closes the loop: Prometheus KPIs → anomaly and SLA checks → LLM remediation plan → scale, restart, rollback or config patch.
- Graded runs on deterministic ground truth instead of how outputs read: built an intent benchmark over a complexity × load × scenario grid, scored resource plans against a formula-based oracle, and compared against different baselines
- 88.9% intent-to-deployment with zero human interventions. The multi-agent pipeline cut config errors from 86.9% (single-LLM baseline) to 0% and remediated 30/30 injected faults.
- Benchmarked 8 open-weight LLMs (3B–120B) as the harness backbone with paired, Bonferroni-corrected tests: accuracy jumps between 7B and 14B (0.38 → 0.69), then plateaus.
AI Development Intern
Nov 2025 – Mar 2026QANTRA
- Curated and validated an Arabic Q&A benchmark to evaluate a RAG pipeline whose keyword-filtering stage reduced retrieval noise (paper in preparation).
- Shipped an OCR + LLM invoice-extraction service.
AI Research Intern
Jul 2025 – Sep 2025Siine
- Benchmarked BKT, DKT and Transformer knowledge-tracing models, lifting recommendation accuracy from 78% to 96%.
- Ran attention and sequence-modeling ablations and integrated learner embeddings into the recommender to reduce cold-start issues; authored the technical report.
Geo AI Research Intern
Sep 2024 – Nov 2024Algerian Space Agency
- Deep learning on satellite imagery for land-cover classification and change detection.
Earlier
- Full-Stack Intern · FikraTek Solutions2024
- Machine Learning Intern · Cellula Technologies2024
- Front-End Developer & UI/UX Designer · Gostu DZ2023 – 2024
Projects
RL for LLM Efficiency
2024Cast three LLM-efficiency problems as RL tasks (prompt-token pruning, attention-head pruning, per-layer pruning + quantization) with multi-objective rewards trading quality (perplexity, output similarity) for latency, memory and FLOPs. On TinyLlama-1.1B, PPO reached 2.09% sparsity (1.100B → 1.077B non-zero weights).
- PyTorch
- Transformers
- Gymnasium
- PPO
HPC: CUDA, OpenMP, MPI
2025CUDA kernels (stencil, dot product, parallel scan) with 101–141× speedups over CPU, each verified against CPU results, plus shared-memory parallelism with OpenMP and distributed regression with MPI.
- CUDA
- C
- OpenMP
- MPI
Other projects
- WLAN intrusion detection · Classification and anomaly-detection models for wireless network attacks.2025
- ECG heartbeat classification · Machine learning and deep learning models to classify ECG heartbeats.2025
- Corporate financial health app · Web app for company budget management with an NLP chatbot.2024
- Gifty · Flutter app connecting gift and flower sellers with buyers.2024
- Algerian forest mapping · Satellite imagery and GIS to monitor forest coverage over time.2023
Core Competencies
Agentic AI & Multi-Agent Systems
LangGraph · multi-agent orchestration · intent-driven automation · tool & function calling · human-in-the-loop checkpoints · shared agent state & memory
LLMs, NLP & Retrieval
LoRA fine-tuning · SFT dataset curation & annotation · RAG pipelines & retrieval evaluation · embeddings & semantic search · prompt engineering & prompt compression · Arabic NLP · OCR + LLM document extraction
Reinforcement Learning, Efficient Inference & HPC
PPO · custom Gymnasium environments · multi-objective reward shaping · token, attention-head & layer pruning · mixed-precision quantization · local LLM serving (vLLM, Ollama) · CUDA kernels · OpenMP · MPI
Cloud-Native Networking & Infrastructure
5G Core (OpenAirInterface) · Kubernetes · Helm · Docker · Prometheus · Grafana · KPI monitoring, anomaly & SLA detection · wireless intrusion detection · microservices & API gateways (Spring Boot) · Linux · GCP
Machine Learning, Data & Developer Stack
Python (primary) · C/C++ · Java · SQL · JavaScript · PyTorch · Hugging Face Transformers · TensorFlow · Scikit-learn · knowledge tracing & recommender systems · time-series analysis · satellite imagery & remote sensing · Spark · Hadoop · Hive · pytest · Git · Flask · Django · React · Next.js · Flutter
Languages
English (C1, IELTS) · French (C2, TCF) · Arabic (native) · Tamazight (native)
Education
MSc, Artificial Intelligence and Data Science
2021 – 2026National School of Artificial Intelligence (ENSIA) · Algiers, Algeria
Thesis with EURECOM: Agentic AI for Orchestration and Management of Telco Cloud. Defended June 2026 with a grade of 18/20 (highest honors).
Selected coursework
- AI & ML
- Machine Learning, Deep Learning, Natural Language Processing, Computer Vision, Reinforcement Learning, Speech Processing, Introduction to Mobile Robotics, AI and Ethics
- Data
- Data Mining, Time Series Analysis and Classification, Big Data Analytics and Visualization, Advanced Databases
- Networks & systems
- Networks and Protocols, Wireless Communication Networks and Systems, Computer and Network Security, High Performance Computing, Operating Systems, Computer Architecture
- Math & optimization
- Statistical Inference, Stochastic Modelling and Simulation, Numerical Methods and Optimisation, Operations Research
- Software
- Software Engineering, Enterprise Computing, Mobile Development, Human Computer Interaction
AI Saturdays
2023World Learning Algeria · Algiers, Algeria
Six-month program in machine learning and deep learning with Python.
Certifications
- Fundamentals of Deep Learning · NVIDIA Deep Learning Institute2022
Community
- Mentor and speaker · GDSC AiFest'24 datathon2024
- Organizing committee · AI for Africa'23 conference2023