Yassir Salmi
AI and Software Engineering at Oracle Labs
Work Experience
I work as an AI Engineer at Oracle Labs, focusing on the evaluation and reliability of LLM-powered agents. My work includes contributing to the development of agentic applications and evaluation for AI systems.
As an AI Engineering Intern, I worked on the performance and deployment side of large language models. I benchmarked inference engines across a range of open-source models and hardware configurations, profiled serving systems to identify bottlenecks, and evaluated optimization techniques that significantly improved latency and throughput. The work helped guide production infrastructure decisions and deepen the team's understanding of efficient LLM serving.
I joined Oracle Labs as a Software Engineering Intern on Data Studio, an internal enterprise data science platform. My main role was integrating an AI assistant that allowed users to interact with data through natural language.
Open Source
NL2Analytics-View SQL Benchmark
GitHubContributed to Oracle's official public benchmark for evaluating LLM and agentic NL2SQL systems on generating Oracle Analytic View SQL. Contributed Analytic View schema/DDL artifacts, gold NL-to-SQL question sets, and agent evaluation specs comparing AV-specialized vs. vanilla agents.
Oracle Database, Analytic Views, Open Agent Spec, NL2SQL
Research
Experimental Evaluation and Benchmarking of Large Language Model Inference Engines (vLLM, llama.cpp, SGLang, etc.)
Master's thesis conducted at Oracle Labs investigating the comparative performance of LLM inference engines. Benchmarked +5 LLM inference engines (vLLM, SGLang, ONNX Runtime GenAI, llama.cpp, etc.) on open source models like Llama, Mistral, phi across CPU and GPU hardware, evaluating latency, throughput, and memory utilization under real-world workloads (summarization, RAG, NER, Q&A, ...). Profiled inference pipelines using NVIDIA Nsight Systems and Linux perf/flame graphs to identify kernel-level bottlenecks. Findings informed production infrastructure decisions on inference engine selection and prompt engineering strategies.
LLMs, vLLM, ONNX Runtime GenAI, SGLang, llama.cpp, inference benchmarking, PagedAttention, prefix caching, GPU/CPU profiling, suffix decoding, speculative decoding
Projects
RAG-Based Document Q&A System
GitHubBuilt a RAG pipeline for Q&A over PDFs with pluggable embeddings/LLMs, incremental ingestion, and RAGAS-based evaluation.
Python, FastAPI, React, LangChain, ChromaDB
Paper Recommender Agent
GitHubBuilt a Plan-and-Execute LLM agent for research paper discovery and structured deep-dive analysis across arXiv, Hugging Face, and Semantic Scholar.
Python, LangChain, Pydantic v2, arXiv API, pdfplumber, Hugging Face Hub
Skills
| Programming | Python, Java, C, C++ |
| Machine Learning | Data Cleaning and Preprocessing, Data Visualization, Regression, Classification, Clustering, Neural Networks (CNN, RNN, LSTM, GRU) |
| Tools | Shell, Git, Docker/Podman, VS Code, IntelliJ IDEA, Jupyter Notebook, Agile Development |
| Languages | English (Proficient), French (Proficient), Arabic (Native) |